Popular in Social Sciences

Books

City Time: On Being Sentenced to Rikers Island

Independent Scholar David A. Campbell

Culture and Society

The City as a Growth Machine: Toward a Political Economy of Place

New York University Harvey Moltoch

Economics

Gender and Economics

University of California Santa Barbara Shelly Lundberg

American Politics

Keeping the Republic

Boston College Marc Landy

Banks and Policy Organisations

Greening the financial sector

European Investment Bank Matteo Ferrazzi

British Politics

Entrepreneurial thinking: the politics and practice of policy impact

University of Sheffield Matthew Flinders

Culture and Society

Carceral Entrapments: Views from the Prison/Street Interface in India

Indian Institute of Technology Delhi Mahuya Bandyopadhya

AI Generated Inventions – Implications for the patent system

View more insights in Social Sciences - Economics - Microeconomics

View more figures from Brandeis University

View the page of Adam Jaffe

Read the Study

Adam Jaffe discusses issues raised for patent processes and policy created by inventions generated by artificial intelligence (AI).

Image courtesy of interviewee. June 17, 2024

Generating Workspace

Export as PDF

This video is a transcript from a discussion about the implications of AI-generated innovations on the patent system. It addresses the legal status of inventions created by AI, noting that patent offices worldwide currently require human inventors to be listed on applications. The conversation touches upon the challenges of distinguishing between human and AI-generated ideas and the potential consequences for patentability standards, particularly regarding the non-obviousness requirement.

The speakers express concerns about the patent office potentially being overwhelmed with applications generated by AI, leading to difficulties in processing. They propose an alternative idea where inventors could certify their AI machines, allowing the patent office to grant patents for applications generated from those certified machines automatically. This video system could streamline patent approval while establishing guidelines and limitations to prevent gaming the system.

The discussion reiterates the need for ongoing research to understand AI's capabilities and explore potential changes in the patent system to ensure it remains functional in an increasingly AI-driven innovation landscape. The speakers emphasize the importance of transparency regarding AI's role in generating inventions, as understanding this will shape future legal frameworks and policies.

Language:

Export as PDF

X-TIMESTAMP-MAP:LOCAL=00:00:00.000,MPEGTS=900000

00:00.025 --> 00:03.895
There is some ambiguity as to exactly what people mean by

00:04.075 --> 00:07.615
and what's going on with AI generated inventions.

00:07.675 --> 00:12.495
But the basic idea is that we now have algorithms,

00:12.685 --> 00:15.375
both general models like chat, GPT

00:15.555 --> 00:18.095
and more specific models that people have developed

00:18.865 --> 00:21.445
for the purpose of searching for inventions

00:22.255 --> 00:26.885
which can produce, ideas, for example,

00:26.985 --> 00:31.685
for new drugs or new devices, which appear to be

00:32.465 --> 00:35.125
in many ways indistinguishable from inventions

00:35.125 --> 00:37.645
that humans would come up with by trying

00:37.645 --> 00:41.405
to work on the same problems at the moment.

00:41.905 --> 00:45.525
Uh, the legal status of those inventions is, when,

00:45.525 --> 00:47.205
when this first started to happen,

00:47.955 --> 00:51.045
some people tried submitting patent applications

00:51.315 --> 00:55.645
with the AI algorithm as the listed invention in, excuse me,

00:55.745 --> 00:58.645
listed inventor for the patent.

00:59.385 --> 01:00.685
And basically patent offices

01:00.785 --> 01:02.605
around the world said, get outta here.

01:03.345 --> 01:05.685
Uh, inventions have to come from people.

01:06.185 --> 01:10.525
Uh, and so the current legal status is frankly,

01:10.585 --> 01:14.045
people are using these algorithms to generate inventions,

01:14.985 --> 01:18.045
and then if they do apply for a patent application,

01:18.235 --> 01:19.845
they list one

01:19.845 --> 01:22.645
or more humans that worked with the algorithm

01:23.385 --> 01:26.525
as the inventions as, excuse me, as the inventors

01:26.945 --> 01:28.845
and on the patent application.

01:29.505 --> 01:32.405
And in that way they're, they're, they're treating the

01:33.555 --> 01:36.045
algorithm that really generated the invention

01:36.825 --> 01:39.405
in the same way they would treat a, a spreadsheet

01:39.465 --> 01:42.805
or a, a, some other software that they had used as part

01:42.805 --> 01:47.205
of their research process, in, in,

01:47.265 --> 01:48.925
in generating the idea for the invention.

01:49.285 --> 01:52.085
I think that the issue here is that,

01:56.395 --> 02:01.055
by, by pretending that the, the algorithm,

02:01.835 --> 02:03.695
cannot generate an invention

02:04.155 --> 02:06.735
and is just sort of like a slide rule

02:06.735 --> 02:09.535
that's helping the inventor, we're kind

02:09.535 --> 02:12.655
of confusing the process of what's really going on.

02:13.195 --> 02:17.175
And it's a little bit, dicey

02:17.175 --> 02:21.975
because under patent law, at least in the us, any,

02:22.355 --> 02:23.975

02:25.575 --> 02:28.285
misleading information in the patent application

02:28.945 --> 02:30.845
can be grounds for invalidating the patent.

02:31.425 --> 02:34.725
Uh, so in some sense we're, we're playing a game

02:34.795 --> 02:37.005
with the patent system by pretending

02:37.195 --> 02:40.965
that the algorithms aren't actually inventing things.

02:41.825 --> 02:44.285
Uh, that is, that is dangerous

02:44.285 --> 02:45.365
because at some point,

02:45.505 --> 02:48.045
if an important invention comes from an algorithm

02:49.185 --> 02:50.885
and it can really be shown

02:51.035 --> 02:53.245
that the algorithm really did invent it,

02:53.265 --> 02:56.605
and they sort of fuzz that over by applying,

02:57.065 --> 02:58.805
by putting a person's name on the

02:58.845 --> 03:01.845
Application who pretty much just took the output

03:01.845 --> 03:03.845
of the algorithm and copied it over

03:03.845 --> 03:05.085
to the patent application,

03:05.715 --> 03:07.685
that could undermine the validity of the patent.

03:07.985 --> 03:10.845
So I think it would be more desirable from a normative

03:10.845 --> 03:13.245
perspective to be more upfront

03:13.305 --> 03:15.285
and straightforward about what's going on here

03:16.185 --> 03:19.365
and say, yes, a human has to be involved in the process.

03:19.905 --> 03:21.405
Uh, that's not a problem,

03:22.265 --> 03:25.405
but we shouldn't pretend that,

03:25.655 --> 03:29.125
these algorithms are not capable of, of actually inventing.

03:29.195 --> 03:31.885
When you apply for a patent, you basically have

03:31.885 --> 03:34.525
to demonstrate two things to be granted a patent.

03:35.225 --> 03:38.925
One is that the idea is new, no one has ever done it before,

03:39.825 --> 03:41.245
but new isn't good enough

03:41.245 --> 03:45.175
because the patent system has the notion that if

03:45.175 --> 03:47.655
what you're proposing hasn't been done before,

03:48.585 --> 03:50.165
but it's sort of trivially obvious,

03:50.235 --> 03:53.125
anybody could have done it if they had ever wanted to,

03:54.155 --> 03:56.645
then you're not entitled to a, to a patent.

03:57.105 --> 04:01.085
Uh, and the way this standard is operationalized,

04:01.415 --> 04:05.605
under US law other countries have related standards

04:06.425 --> 04:10.245
is that what the examiner at the patent office does is ask

04:10.245 --> 04:14.605
the question whether a knowledgeable person who is familiar

04:14.605 --> 04:17.285
with this technology could have come up

04:17.285 --> 04:19.365
with this invention if someone had asked them to.

04:20.105 --> 04:21.845
And if the answer to that question is yes,

04:22.125 --> 04:24.165
a knowledgeable person in this area could have come up

04:24.165 --> 04:26.725
with this invention, even though it turns out nobody ever

04:26.725 --> 04:30.845
did 'cause nobody cared, then you can't get a patent even if

04:30.845 --> 04:32.485
what you're doing hasn't been done before.

04:32.545 --> 04:34.485
And that's called the obviousness standard

04:34.585 --> 04:37.365
or the non obviousness standard.

04:38.715 --> 04:40.925
Once we introduce the reality, and,

04:40.945 --> 04:43.045
and this really would be a problem, regardless

04:43.225 --> 04:46.085
of whether we explicitly recognize the machine

04:46.305 --> 04:48.085
as the inventor or,

04:48.585 --> 04:51.645
or whether we continue with the current regime of sort

04:51.645 --> 04:54.205
of fuzzing over the role of the machine, this problem was,

04:54.265 --> 04:57.645
is there either way, we now have to ask the question.

04:58.195 --> 05:01.205
When we try to figure out whether an invention is obvious,

05:02.235 --> 05:05.505
would we say, well, an algorithm could have come up with it,

05:05.735 --> 05:08.345
then it's obvious and it can't be patented.

05:09.165 --> 05:10.345
That's kind of a problem

05:10.345 --> 05:13.225
because in principle, these algorithms could come up

05:13.225 --> 05:14.225
with almost anything.

05:15.085 --> 05:18.785
And so if we now articulate in the obviousness standard

05:19.055 --> 05:21.665
that if an algorithm could have come up with it,

05:21.775 --> 05:23.745
then it's obvious and not patentable,

05:24.445 --> 05:25.825
we might enter a world in which

05:26.485 --> 05:28.265
either nothing is patentable,

05:28.265 --> 05:29.945
or at the very least, it's very hard

05:29.945 --> 05:32.785
to figure out what's patentable and what's not.

05:33.445 --> 05:37.265
So I think it may seem a little bit, inconsistent,

05:37.445 --> 05:38.585
but I don't think it is.

05:38.685 --> 05:41.065
On the one hand, I'm arguing we should recognize

05:41.575 --> 05:44.625
that these machines really do create inventions,

05:45.405 --> 05:47.265
but at the same time, the standard

05:47.365 --> 05:50.425
for patentability should still be human inventors.

05:51.205 --> 05:53.945
And we should ask the question, could a human

05:54.445 --> 05:58.245
who is knowledgeable in this technology have come up

05:58.245 --> 06:00.245
with this idea if someone had asked them to?

06:00.905 --> 06:04.205
If the answer to that is no, not, not, not necessarily,

06:04.795 --> 06:07.405
then it's patentable, whether it came from a machine

06:07.425 --> 06:08.565
or it came from a person.

06:08.785 --> 06:11.845
So one of the things that you, you might say, if you

06:12.395 --> 06:15.445
sort of came as far as we've come so far as you might say,

06:15.565 --> 06:18.605
okay, you're right, it makes sense to acknowledge

06:18.605 --> 06:20.725
that machines can generate these inventions,

06:21.265 --> 06:22.605
but what we ought to do is we

06:22.605 --> 06:23.685
ought to treat them differently.

06:23.785 --> 06:26.765
We ought to have a different set of examination

06:27.025 --> 06:29.645
and granting standards for inventions

06:29.645 --> 06:30.885
that come from machines.

06:31.305 --> 06:34.605
And that's actually a pretty reasonable argument on some

06:34.605 --> 06:38.005
level, because from a economic perspective,

06:39.185 --> 06:43.755
the cheaper it is to make an invention, the less you need

06:43.815 --> 06:47.635
to give the person you've granted a patent, a very powerful

06:48.285 --> 06:50.315
right, to, to incentivize them.

06:50.315 --> 06:52.955
Because if it's very easy to do, you don't need

06:53.075 --> 06:55.915
to be giving them this very powerful patent in return.

06:56.495 --> 06:58.195
So it, you could easily argue

06:58.425 --> 07:02.715
that we should let the machines in effect apply for patents,

07:03.095 --> 07:04.515
but we should treat them differently.

07:04.615 --> 07:06.315
We should have a different set of standards

07:06.315 --> 07:08.635
and a different set of, rights

07:08.705 --> 07:12.235
that come along within a patent, that's granted

07:12.455 --> 07:14.395
for an invention that comes from a machine.

07:14.935 --> 07:17.635
And as I say, that's not a crazy argument that,

07:17.635 --> 07:19.315
that there's reasons to think that way.

07:19.815 --> 07:22.555
The problem with that argument is a very practical one,

07:23.085 --> 07:26.755
which is when you look at the invention itself, the thing

07:26.755 --> 07:27.995
that has been created,

07:29.025 --> 07:31.525
you can't really tell whether it came from a person

07:31.865 --> 07:33.125
or a, or a machine.

07:33.985 --> 07:36.445
And so if you have a different set of rules

07:36.625 --> 07:38.245
for those two different categories,

07:38.865 --> 07:41.725
you're gonna create incentives once again for people

07:41.905 --> 07:44.565
to obfuscate what really happened

07:45.145 --> 07:48.725
and to make it seem like the role of the algorithm

07:48.865 --> 07:52.085
or the machine was less than it really was, so

07:52.085 --> 07:56.085
that they can qualify for this other kind of patent.

07:56.345 --> 07:58.285
And we have some experience with that.

07:58.385 --> 08:01.445
For example, in the, in around the turn

08:01.445 --> 08:05.445
of the 21st century when business method patents became a

08:05.455 --> 08:09.805
major category of patents, the patent office was worried

08:09.915 --> 08:13.205
that it was granting too many business method patents.

08:13.225 --> 08:16.245
So it created a special extra review

08:17.205 --> 08:19.465
for business method patent applications.

08:19.875 --> 08:22.065
Guess what happened? All the most

08:22.065 --> 08:25.265
of the business method applications disappeared,

08:25.365 --> 08:26.785
and people figured out ways

08:27.405 --> 08:29.585
to make their application look like it

08:29.585 --> 08:30.665
wasn't a business method.

08:31.045 --> 08:33.705
So they wouldn't qual they wouldn't be forced into

08:33.705 --> 08:34.865
this other track.

08:35.005 --> 08:37.705
And I think it's just at a practical level,

08:37.805 --> 08:39.985
if you have different rules for different kinds

08:39.985 --> 08:43.065
of applications, what's gonna happen is you're gonna have a

08:43.065 --> 08:45.985
lot of gaming and a lot of trickery going on.

08:46.005 --> 08:49.865
So I think just as a practical matter, it's not a good idea

08:50.245 --> 08:53.625
to have different rules for AI patents, even though

08:54.165 --> 08:56.285
at a theoretical level, I understand

08:56.395 --> 08:58.845
that there is a decent argument for doing that.

08:59.125 --> 09:01.325
I don't pretend to know the solutions,

09:01.505 --> 09:03.125
but I think we should start talking

09:03.125 --> 09:04.285
about some possibilities.

09:04.505 --> 09:08.725
So, I, I and my co-authors have thrown out an idea.

09:08.945 --> 09:10.205
I'm not sure it's a great idea,

09:10.225 --> 09:12.285
but it's a way to try to stimulate some

09:12.285 --> 09:13.565
discussion around this.

09:14.345 --> 09:15.965
And the issue is the following.

09:16.545 --> 09:19.245
Uh, once we have these algorithms

09:19.245 --> 09:22.165
that are very good at generating inventions,

09:22.795 --> 09:26.245
there's a big danger that we're, the patent offices

09:26.345 --> 09:29.885
of the world are gonna be overwhelmed with thousands

09:29.985 --> 09:33.085
and thousands of applications from these algorithms,

09:33.495 --> 09:35.885
which is gonna just make it very difficult for them

09:36.065 --> 09:38.445
to process them all and, and keep up.

09:39.185 --> 09:41.405
Now, you could say, well, they can use AI

09:42.065 --> 09:45.245
to evaluate the applications in the same way

09:45.245 --> 09:49.205
that the inventors are using AI to generate the inventions.

09:49.665 --> 09:51.045
And there's something to that.

09:51.825 --> 09:55.685
But I suspect that there's, there's two problems with that.

09:56.185 --> 09:59.365
One is practical that the government just will probably

09:59.365 --> 10:01.445
always be behind the private sector

10:01.625 --> 10:02.845
in developing these methods.

10:03.385 --> 10:05.045
And the second is sort of political.

10:05.205 --> 10:08.965
I, I'm not sure we could go to a system where we had

10:09.535 --> 10:12.925
algorithms deciding which patents the government is granting

10:13.345 --> 10:15.885
and have people just people be comfortable with that.

10:16.665 --> 10:21.245
So what we propose as an alternative is to say, well, look,

10:21.945 --> 10:25.645
what's going on is someone's training an algorithm based

10:25.645 --> 10:29.005
on previous patents to generate ideas

10:29.275 --> 10:33.325
that meet the standards of patentability, that are novel

10:33.465 --> 10:34.525
and that are not obvious.

10:35.595 --> 10:39.395
Once they've done that, what we could do is we could say

10:39.705 --> 10:42.675
that if you have created such an algorithm

10:42.675 --> 10:46.795
or such a machine, you can submit your machine

10:46.935 --> 10:47.955
to the patent office.

10:48.135 --> 10:51.155
So this is not an application for a specific patent,

10:51.745 --> 10:54.915
it's an application for approval of your machine

10:55.335 --> 10:58.755
as one which generates proposals

10:58.755 --> 11:00.755
for patents which meet the standards.

11:01.375 --> 11:03.675
And the patent office could review the machine.

11:04.575 --> 11:06.155
And if it really does do that,

11:06.305 --> 11:08.755
they could in some sense certify that machine

11:09.535 --> 11:11.875
and then any application that came from

11:11.875 --> 11:13.635
that machine would be granted,

11:13.815 --> 11:16.035
and the patent office wouldn't have to review it.

11:16.705 --> 11:20.275
They could continue to review, say a random sample, a few

11:20.295 --> 11:23.555
of the proposals from that machine every year to make sure

11:24.145 --> 11:25.875
that it's still good,

11:26.335 --> 11:28.475
but they wouldn't have to review all of them.

11:29.715 --> 11:32.735
Now, this has a couple of potential benefits.

11:32.955 --> 11:35.215
One is it just makes it more practical

11:35.215 --> 11:37.055
for the patent office to function.

11:37.835 --> 11:41.295
But the other is, it solves this other problem we were

11:41.295 --> 11:44.055
talking about a second ago, which is

11:45.295 --> 11:46.305
this would be something

11:46.305 --> 11:48.985
that the applicant would have an incentive to do

11:48.985 --> 11:51.865
because it makes it easier for them to get patents.

11:52.525 --> 11:55.645
So you could say, okay, you can submit your machine

11:56.185 --> 11:58.245
and if it's valid, we'll certify it,

11:58.625 --> 11:59.845
and then you're home free.

12:00.065 --> 12:02.005
All of your applications are granted,

12:02.265 --> 12:03.885
but they're going

12:03.885 --> 12:05.925
to have different rights than other patents.

12:06.035 --> 12:09.045
They're gonna maybe be shorter or more expensive,

12:09.665 --> 12:12.965
or you're gonna have obligations to license those patents

12:12.965 --> 12:15.525
to people who request it to limit the power

12:15.555 --> 12:16.605
that you've been granted.

12:17.625 --> 12:20.725
And that's a way of having a different kind of patents

12:20.905 --> 12:24.365
for AI inventions that doesn't have this problem

12:24.625 --> 12:25.645
of inducing gaming

12:25.645 --> 12:29.125
because it's up to the applicant to decide that they want

12:29.125 --> 12:30.125
to go this route.

12:30.545 --> 12:32.525
And if they don't wanna go this route, that's fine.

12:32.525 --> 12:34.245
They don't have to submit their machine.

12:34.245 --> 12:37.525
They could continue to submit each patent the machine

12:37.845 --> 12:39.365
generates through the normal route,

12:40.025 --> 12:42.445
and they would have to bear the same delays

12:42.505 --> 12:44.605
and the same,

12:44.985 --> 12:46.005
but then they would get the same

12:46.025 --> 12:47.405
rights that other people get.

12:47.585 --> 12:50.965
So we think it's worth thinking about this kind of, sort

12:50.965 --> 12:55.085
of alternative way of looking at AI patents that deal

12:55.085 --> 12:59.245
with the reality without hopefully creating new,

12:59.985 --> 13:01.765
gaming problems in in the system.

13:01.865 --> 13:04.925
Anyone who's proposing solutions needs to acknowledge that

13:05.585 --> 13:07.525
it could turn out to be a bad idea

13:07.525 --> 13:09.925
because the technology might evolve in a way

13:09.945 --> 13:11.085
we don't anticipate.

13:11.745 --> 13:15.205
Uh, I think the other problem that this idea

13:16.725 --> 13:20.625
doesn't deal with, and perhaps exacerbates, is

13:20.625 --> 13:24.785
that we already have an issue that the ownership

13:24.785 --> 13:29.655
of patents generally has become in the last couple decades,

13:30.365 --> 13:34.015
more and more concentrated in a relatively small number

13:34.075 --> 13:36.375
of big players, big companies.

13:37.555 --> 13:39.815
And that's not good for the economy,

13:40.355 --> 13:44.095
for the innovation system, or frankly for society.

13:44.635 --> 13:48.695
Uh, and, AI as a technology

13:49.515 --> 13:52.255
is already likely to exacerbate this problem.

13:53.435 --> 13:57.015
And by creating this mechanism that makes it in some sense,

13:57.205 --> 14:00.215
easy for someone who's created an algorithm

14:00.835 --> 14:02.975
to get a lot more patents,

14:03.995 --> 14:05.895
we we might be exacerbating that problem.

14:06.235 --> 14:08.615
Um, and I think that's something to think about.

14:09.615 --> 14:13.175
I think there are, there are deep questions about whether

14:13.755 --> 14:18.215
if this plays out in the most successful way

14:18.215 --> 14:21.015
that some people are anticipating, it might.

14:22.655 --> 14:25.695
I think there's a question of whether the patent system,

14:26.075 --> 14:29.895
can really survive at all in, in the long run, in the face

14:29.955 --> 14:32.895
of a technology which can so easily generate

14:33.745 --> 14:34.855
valid applications.

14:35.635 --> 14:38.615
Um, so, you know, those are questions we,

14:38.795 --> 14:40.375
we don't know the answer to.

14:40.715 --> 14:45.125
Um, but I just think that, what we're trying

14:45.125 --> 14:47.165
to do here is stimulate some discussion

14:47.465 --> 14:48.765
around these possibilities.

14:49.785 --> 14:52.165
Um, the one thing we haven't talked about

14:52.265 --> 14:56.415
so far in this discussion is when people think about

14:56.685 --> 15:00.095
patents, they think about it as an incentive

15:00.195 --> 15:01.535
for people to invent things.

15:02.875 --> 15:04.415
Um, but which is right.

15:04.515 --> 15:06.015
And that's, there's this language, for example,

15:06.015 --> 15:08.735
in the US Constitution about encouraging the

15:08.935 --> 15:10.015
creative arts and so forth.

15:10.755 --> 15:13.775
But patents have another important function that,

15:13.805 --> 15:16.335
that is underappreciated, which is

15:16.335 --> 15:19.175
that even if people are inventing things right

15:19.175 --> 15:22.425
and left, we don't really get the benefits

15:22.425 --> 15:23.425
of those inventions

15:23.845 --> 15:28.465
unless they also make investments that are necessary

15:28.645 --> 15:32.665
to carry the invention from the idea stage to something

15:32.665 --> 15:34.625
that's really valuable to society.

15:34.765 --> 15:38.465
So the, the clearest example of this is a scientist

15:38.765 --> 15:42.185
or an algorithm can come up with an idea for a new drug,

15:42.305 --> 15:44.425
a new chemical molecule, which based on

15:45.095 --> 15:48.745
what we know about other molecules we think might,

15:49.205 --> 15:50.585
you know, kill cancer.

15:51.365 --> 15:54.865
Um, but that's not worth anything until you spend the money

15:55.045 --> 15:58.305
to test it in clinical trials and prove that it works

15:58.365 --> 16:02.035
and that it's safe, that process costs, you know, hundreds

16:02.035 --> 16:03.435
of millions to a billion dollars.

16:04.415 --> 16:08.835
No one's going to do that if they don't know that.

16:08.975 --> 16:12.235
If it does turn out to be valuable, they will be able to,

16:12.855 --> 16:14.555
market it and make money from it.

16:14.735 --> 16:17.315
So even if we get to the point where

16:17.315 --> 16:21.055
because of ai, we've got more inventions than we know what

16:21.055 --> 16:24.965
to do with, we still need some kind of mechanism

16:26.125 --> 16:27.865
to allow people to

16:29.035 --> 16:33.285
make the investments in commercializing those ideas, knowing

16:33.355 --> 16:36.125
that they will have some kind of protected market

16:36.825 --> 16:38.965
for selling the product if it pans out.

16:39.905 --> 16:43.405
And what's interesting about this is,

16:44.715 --> 16:48.695
we have for now 600 years, we've tied

16:49.575 --> 16:52.055
granting that kind of market protection

16:52.915 --> 16:54.935
to being the first one to come up with an idea.

16:56.325 --> 16:57.865
But on some deep level,

16:58.795 --> 17:01.165
that connection is not really necessary.

17:01.505 --> 17:06.005
You could imagine a different system in which we sort of

17:06.515 --> 17:09.485
acknowledge that coming up with the new ideas, you know,

17:09.485 --> 17:11.085
that's not really an issue anymore.

17:11.345 --> 17:13.725
So we're not gonna reward people for that,

17:14.425 --> 17:16.565
but we're still going to grant people

17:17.535 --> 17:18.765
under certain conditions

17:19.105 --> 17:22.965
and in return for certain other obligations, well,

17:22.985 --> 17:25.085
you might call an exclusive marketing, right?

17:26.465 --> 17:28.525
So you could go to the government and,

17:28.625 --> 17:30.045
and I'm making this up now,

17:30.185 --> 17:33.405
but suppose you could go to the government, you could pay

17:33.945 --> 17:37.885
$10 million or $50 million, some, some significant amount

17:37.885 --> 17:39.525
of money, and you could say, look,

17:39.725 --> 17:41.285
I didn't, I didn't invent this.

17:41.485 --> 17:44.245
I didn't come up with it. An algorithm came up with it,

17:44.665 --> 17:47.725
or Sam from faculty came up with it, but,

17:48.105 --> 17:50.335
but nobody is, is developing it.

17:50.495 --> 17:52.935
Nobody's spending the money to convert it to a product.

17:53.595 --> 17:55.735
So I will pay a big chunk

17:55.735 --> 17:57.975
of money if the government will grant me

17:58.595 --> 18:01.015
an exclusive marketing right, for this idea.

18:02.035 --> 18:06.415
And the only condition would be no one else has stepped

18:06.415 --> 18:08.255
forward and said they're gonna do it.

18:08.955 --> 18:11.535
Uh, and I am gonna meet some conditions.

18:11.715 --> 18:15.335
And so again, I think this is the kind of idea that we need

18:15.335 --> 18:19.335
to be thinking about as we move into a world in which the

18:19.335 --> 18:20.815
whole process of invention

18:20.915 --> 18:22.335
and innovation could be very

18:22.335 --> 18:23.375
different from what we're used to.

18:23.655 --> 18:26.255
I think we need to continue to research

18:26.315 --> 18:30.295
and people are already working on this to understand better

18:31.435 --> 18:35.165
what AI really can do and will be able to do.

18:35.525 --> 18:37.085
I mean, there's some people who believe

18:37.085 --> 18:38.965
that the time will come where it can do any,

18:38.965 --> 18:40.045
everything people can do.

18:40.675 --> 18:44.085
There's some people who believe that it will always be

18:44.665 --> 18:46.525
in a way complimentary

18:46.625 --> 18:49.125
to human intelligence rather than a complete

18:49.175 --> 18:50.405
substitute for it.

18:51.625 --> 18:54.525
And the answer to that makes a big difference in terms of,

18:54.945 --> 18:56.925
you know, how we want to treat this stuff.

18:57.225 --> 18:59.165
And I think one of the reasons

18:59.725 --> 19:04.045
I believe the current legal regime, which sort of encourages

19:04.755 --> 19:09.175
obfuscation of the role of the technology is unfortunate is

19:09.175 --> 19:12.775
because understanding exactly how the the AI is working

19:13.595 --> 19:17.135
is key to deciding in the future how we want to treat it.

19:17.595 --> 19:20.415
And if we're encouraging people to sort of cover up

19:20.445 --> 19:23.575
what the AI is actually doing, it just makes it harder

19:24.195 --> 19:26.375
to understand and therefore to deal with it.

00:00 - There is some ambiguity as to exactly what people mean by and what's going on with AI generated inventions. But the basic idea is that we now have algorithms, both general models like chat, GPT and more specific models that people have developed for the purpose of searching for inventions

00:22 - which can produce, ideas, for example, for new drugs or new devices, which appear to be in many ways indistinguishable from inventions that humans would come up with by trying to work on the same problems at the moment. the legal status of those inventions is, when, when this first started to happen, some people tried submitting patent applications with the AI algorithm as the listed invention in, excuse me, listed inventor for the patent.

00:59 - And basically patent offices around the world said, get outta here. inventions have to come from people. and so the current legal status is frankly, people are using these algorithms to generate inventions, and then if they do apply for a patent application, they list one or more humans that worked with the algorithm as the inventions as, excuse me, as the inventors and on the patent application.

01:29 - And in that way they're, they're, they're treating the algorithm that really generated the invention in the same way they would treat a, a spreadsheet or a, a, some other software that they had used as part of their research process, in, in, in generating the idea for the invention. I think that the issue here is that, by, by pretending that the, the algorithm, cannot generate an invention and is just sort of like a slide rule

02:06 - that's helping the inventor, we're kind of confusing the process of what's really going on. And it's a little bit, dicey because under patent law, at least in the us, any, misleading information in the patent application can be grounds for invalidating the patent. so in some sense we're, we're playing a game with the patent system by pretending that the algorithms aren't actually inventing things.

02:41 - that is, that is dangerous because at some point, if an important invention comes from an algorithm and it can really be shown that the algorithm really did invent it, and they sort of fuzz that over by applying, by putting a person's name on the Application who pretty much just took the output of the algorithm and copied it over to the patent application,

03:05 - that could undermine the validity of the patent. So I think it would be more desirable from a normative perspective to be more upfront and straightforward about what's going on here and say, yes, a human has to be involved in the process. that's not a problem, but we shouldn't pretend that, these algorithms are not capable of, of actually inventing. When you apply for a patent, you basically have to demonstrate two things to be granted a patent.

03:35 - One is that the idea is new, no one has ever done it before, but new isn't good enough because the patent system has the notion that if what you're proposing hasn't been done before, but it's sort of trivially obvious, anybody could have done it if they had ever wanted to, then you're not entitled to a, to a patent. and the way this standard is operationalized, under US law other countries have related standards is that what the examiner at the patent office does is ask

04:10 - the question whether a knowledgeable person who is familiar with this technology could have come up with this invention if someone had asked them to. And if the answer to that question is yes, a knowledgeable person in this area could have come up with this invention, even though it turns out nobody ever did 'cause nobody cared, then you can't get a patent even if what you're doing hasn't been done before. And that's called the obviousness standard or the non obviousness standard.

04:38 - Once we introduce the reality, and, and this really would be a problem, regardless of whether we explicitly recognize the machine as the inventor or, or whether we continue with the current regime of sort of fuzzing over the role of the machine, this problem was, is there either way, we now have to ask the question. When we try to figure out whether an invention is obvious, would we say, well, an algorithm could have come up with it, then it's obvious and it can't be patented.

05:09 - That's kind of a problem because in principle, these algorithms could come up with almost anything. And so if we now articulate in the obviousness standard that if an algorithm could have come up with it, then it's obvious and not patentable, we might enter a world in which either nothing is patentable, or at the very least, it's very hard to figure out what's patentable and what's not.

05:33 - So I think it may seem a little bit, inconsistent, but I don't think it is. On the one hand, I'm arguing we should recognize that these machines really do create inventions, but at the same time, the standard for patentability should still be human inventors. And we should ask the question, could a human who is knowledgeable in this technology have come up with this idea if someone had asked them to? If the answer to that is no, not, not, not necessarily,

06:04 - then it's patentable, whether it came from a machine or it came from a person. So one of the things that you, you might say, if you sort of came as far as we've come so far as you might say, okay, you're right, it makes sense to acknowledge that machines can generate these inventions, but what we ought to do is we ought to treat them differently. We ought to have a different set of examination and granting standards for inventions

06:29 - that come from machines. And that's actually a pretty reasonable argument on some level, because from a economic perspective, the cheaper it is to make an invention, the less you need to give the person you've granted a patent, a very powerful right, to, to incentivize them. Because if it's very easy to do, you don't need to be giving them this very powerful patent in return. So it, you could easily argue that we should let the machines in effect apply for patents,

07:03 - but we should treat them differently. We should have a different set of standards and a different set of, rights that come along within a patent, that's granted for an invention that comes from a machine. And as I say, that's not a crazy argument that, that there's reasons to think that way. The problem with that argument is a very practical one, which is when you look at the invention itself, the thing that has been created,

07:29 - you can't really tell whether it came from a person or a, or a machine. And so if you have a different set of rules for those two different categories, you're gonna create incentives once again for people to obfuscate what really happened and to make it seem like the role of the algorithm or the machine was less than it really was, so that they can qualify for this other kind of patent. And we have some experience with that.

07:58 - For example, in the, in around the turn of the 21st century when business method patents became a major category of patents, the patent office was worried that it was granting too many business method patents. So it created a special extra review for business method patent applications. Guess what happened? All the most of the business method applications disappeared, and people figured out ways to make their application look like it

08:29 - wasn't a business method. So they wouldn't qual they wouldn't be forced into this other track. And I think it's just at a practical level, if you have different rules for different kinds of applications, what's gonna happen is you're gonna have a lot of gaming and a lot of trickery going on. So I think just as a practical matter, it's not a good idea to have different rules for AI patents, even though at a theoretical level, I understand

08:56 - that there is a decent argument for doing that. I don't pretend to know the solutions, but I think we should start talking about some possibilities. So, I, I and my co-authors have thrown out an idea. I'm not sure it's a great idea, but it's a way to try to stimulate some discussion around this. And the issue is the following. once we have these algorithms

09:19 - that are very good at generating inventions, there's a big danger that we're, the patent offices of the world are gonna be overwhelmed with thousands and thousands of applications from these algorithms, which is gonna just make it very difficult for them to process them all and, and keep up. Now, you could say, well, they can use AI to evaluate the applications in the same way that the inventors are using AI to generate the inventions. And there's something to that.

09:51 - But I suspect that there's, there's two problems with that. One is practical that the government just will probably always be behind the private sector in developing these methods. And the second is sort of political. I, I'm not sure we could go to a system where we had algorithms deciding which patents the government is granting and have people just people be comfortable with that. So what we propose as an alternative is to say, well, look, what's going on is someone's training an algorithm based

10:25 - on previous patents to generate ideas that meet the standards of patentability, that are novel and that are not obvious. Once they've done that, what we could do is we could say that if you have created such an algorithm or such a machine, you can submit your machine to the patent office. So this is not an application for a specific patent, it's an application for approval of your machine as one which generates proposals

10:58 - for patents which meet the standards. And the patent office could review the machine. And if it really does do that, they could in some sense certify that machine and then any application that came from that machine would be granted, and the patent office wouldn't have to review it. They could continue to review, say a random sample, a few of the proposals from that machine every year to make sure that it's still good,

11:26 - but they wouldn't have to review all of them. Now, this has a couple of potential benefits. One is it just makes it more practical for the patent office to function. But the other is, it solves this other problem we were talking about a second ago, which is this would be something that the applicant would have an incentive to do because it makes it easier for them to get patents. So you could say, okay, you can submit your machine

11:56 - and if it's valid, we'll certify it, and then you're home free. All of your applications are granted, but they're going to have different rights than other patents. They're gonna maybe be shorter or more expensive, or you're gonna have obligations to license those patents to people who request it to limit the power that you've been granted. And that's a way of having a different kind of patents

12:20 - for AI inventions that doesn't have this problem of inducing gaming because it's up to the applicant to decide that they want to go this route. And if they don't wanna go this route, that's fine. They don't have to submit their machine. They could continue to submit each patent the machine generates through the normal route, and they would have to bear the same delays and the same,

12:44 - but then they would get the same rights that other people get. So we think it's worth thinking about this kind of, sort of alternative way of looking at AI patents that deal with the reality without hopefully creating new, gaming problems in in the system. Anyone who's proposing solutions needs to acknowledge that it could turn out to be a bad idea because the technology might evolve in a way we don't anticipate.

13:11 - I think the other problem that this idea doesn't deal with, and perhaps exacerbates, is that we already have an issue that the ownership of patents generally has become in the last couple decades, more and more concentrated in a relatively small number of big players, big companies. And that's not good for the economy, for the innovation system, or frankly for society. and, AI as a technology is already likely to exacerbate this problem.

13:53 - And by creating this mechanism that makes it in some sense, easy for someone who's created an algorithm to get a lot more patents, we we might be exacerbating that problem. and I think that's something to think about. I think there are, there are deep questions about whether if this plays out in the most successful way that some people are anticipating, it might. I think there's a question of whether the patent system, can really survive at all in, in the long run, in the face

14:29 - of a technology which can so easily generate valid applications. so, you know, those are questions we, we don't know the answer to. but I just think that, what we're trying to do here is stimulate some discussion around these possibilities. the one thing we haven't talked about so far in this discussion is when people think about patents, they think about it as an incentive

15:00 - for people to invent things. but which is right. And that's, there's this language, for example, in the US Constitution about encouraging the creative arts and so forth. But patents have another important function that, that is underappreciated, which is that even if people are inventing things right and left, we don't really get the benefits of those inventions

15:23 - unless they also make investments that are necessary to carry the invention from the idea stage to something that's really valuable to society. So the, the clearest example of this is a scientist or an algorithm can come up with an idea for a new drug, a new chemical molecule, which based on what we know about other molecules we think might, you know, kill cancer. but that's not worth anything until you spend the money to test it in clinical trials and prove that it works

15:58 - and that it's safe, that process costs, you know, hundreds of millions to a billion dollars. No one's going to do that if they don't know that. If it does turn out to be valuable, they will be able to, market it and make money from it. So even if we get to the point where because of ai, we've got more inventions than we know what to do with, we still need some kind of mechanism to allow people to make the investments in commercializing those ideas, knowing

16:33 - that they will have some kind of protected market for selling the product if it pans out. And what's interesting about this is, we have for now 600 years, we've tied granting that kind of market protection to being the first one to come up with an idea. But on some deep level, that connection is not really necessary. You could imagine a different system in which we sort of acknowledge that coming up with the new ideas, you know,

17:09 - that's not really an issue anymore. So we're not gonna reward people for that, but we're still going to grant people under certain conditions and in return for certain other obligations, well, you might call an exclusive marketing, right? So you could go to the government and, and I'm making this up now, but suppose you could go to the government, you could pay $10 million or $50 million, some, some significant amount

17:37 - of money, and you could say, look, I didn't, I didn't invent this. I didn't come up with it. An algorithm came up with it, or Sam from faculty came up with it, but, but nobody is, is developing it. Nobody's spending the money to convert it to a product. So I will pay a big chunk of money if the government will grant me an exclusive marketing right, for this idea. And the only condition would be no one else has stepped

18:06 - forward and said they're gonna do it. and I am gonna meet some conditions. And so again, I think this is the kind of idea that we need to be thinking about as we move into a world in which the whole process of invention and innovation could be very different from what we're used to. I think we need to continue to research and people are already working on this to understand better what AI really can do and will be able to do.

18:35 - I mean, there's some people who believe that the time will come where it can do any, everything people can do. There's some people who believe that it will always be in a way complimentary to human intelligence rather than a complete substitute for it. And the answer to that makes a big difference in terms of, you know, how we want to treat this stuff. And I think one of the reasons

18:59 - I believe the current legal regime, which sort of encourages obfuscation of the role of the technology is unfortunate is because understanding exactly how the the AI is working is key to deciding in the future how we want to treat it. And if we're encouraging people to sort of cover up what the AI is actually doing, it just makes it harder to understand and therefore to deal with it.

Faculti AI can make mistakes