WEBVTT

00:00:00.000 --> 00:00:03.650
Who will pay for AI?

00:00:03.650 --> 00:00:07.300
By the end, you will see why price and the means to stop AI belong in one decision.

00:00:07.300 --> 00:00:10.950
Some of the big providers now charge by how much you use.

00:00:10.950 --> 00:00:14.600
On the same day, a developer's AI was reported to have caused trouble in systems it

00:00:14.600 --> 00:00:18.250
did not own.

00:00:18.250 --> 00:00:21.900
When a drug company uses such tools, who is responsible for stopping them?

00:00:21.900 --> 00:00:25.550
The user builds in the safeguards and checks that they work.

00:00:25.550 --> 00:00:29.200
So the cost estimate, the stop procedure and the person who checks should be set

00:00:29.200 --> 00:00:32.850
together.

00:00:32.850 --> 00:00:36.500
We will follow the money, then the incidents, and then the plants that make

00:00:36.500 --> 00:00:40.150
medicines.

00:00:40.150 --> 00:00:45.018
Let's start with how money changes hands.

00:00:45.018 --> 00:00:49.886
When the shape of the cost changes, so does the choice of which tasks to give to

00:00:49.886 --> 00:00:54.753
AI.One giant has dropped its flat fee per person and now charges by use.

00:00:54.753 --> 00:00:59.621
At another, the free tier is expected to shrink to the simplest tool.

00:00:59.621 --> 00:01:04.489
I read this as a clear signal.

00:01:04.489 --> 00:01:09.357
The phase of gathering users is over.

00:01:09.357 --> 00:01:14.224
The push to earn back the investment has begun.

00:01:14.224 --> 00:01:19.092
Meanwhile, heavy users are pressing the developers to lower prices.Under pay-per-

00:01:19.092 --> 00:01:23.960
use, the busiest departments get the biggest bills.

00:01:23.960 --> 00:01:28.828
Monthly costs also become hard to predict.

00:01:28.828 --> 00:01:33.696
Firms that fix budgets at the start of the year will feel this most.Many pharma

00:01:33.696 --> 00:01:38.563
teams have begun trying AI for literature summaries and draft materials.

00:01:38.563 --> 00:01:43.431
Good results from the free period cannot simply justify a production budget.

00:01:43.431 --> 00:01:48.299
I would first estimate which tasks will use how much, and then choose the contract.

00:01:48.299 --> 00:01:53.167
The same use can cost very different amounts in a year under different contracts.

00:01:53.167 --> 00:01:58.034
Usage logs also let you compare cost and benefit later.

00:01:58.034 --> 00:02:02.902
Still, why can heavy users demand lower prices?

00:02:02.902 --> 00:02:07.770
Because the developers are fighting each other for customers.

00:02:07.780 --> 00:02:11.791
Prices are going up while companies fight for users.

00:02:11.791 --> 00:02:15.801
These two moves seem to clash.

00:02:15.801 --> 00:02:19.811
So why do they happen at the same time?At first, developers give their tools away or

00:02:19.811 --> 00:02:23.822
sell them cheaply to build a user base.

00:02:23.822 --> 00:02:27.833
Once that base is in place, they move to making money.

00:02:27.833 --> 00:02:31.843
By then, though, the biggest buyers hold a strong position.

00:02:31.843 --> 00:02:35.853
They can switch to a rival, so they hold firm on price.

00:02:35.853 --> 00:02:39.864
Developers want revenue, and heavy users want discounts.

00:02:39.864 --> 00:02:43.875
Their interests collide.In fact, the big players are luring young companies with a

00:02:43.875 --> 00:02:47.885
year of free use.

00:02:47.885 --> 00:02:51.895
A new developer that publishes its model's internals is also being named as a

00:02:51.895 --> 00:02:55.906
challenger.

00:02:55.906 --> 00:02:59.916
One report even asks whether customers can be helped to depend less on a single

00:02:59.916 --> 00:03:03.927
provider.For a drug company, this is leverage in negotiations.

00:03:03.927 --> 00:03:07.938
Avoid relying on one provider, and keep the option to switch.

00:03:07.938 --> 00:03:11.948
Then a price rise is easier to resist.

00:03:11.948 --> 00:03:15.958
But each switch means checking the quality of the output text again.

00:03:15.958 --> 00:03:19.969
That effort belongs in the cost as well.

00:03:19.969 --> 00:03:23.979
While the haggling over price went on, a developer's AI was causing trouble out in

00:03:23.979 --> 00:03:27.990
the world.

00:03:27.980 --> 00:03:32.326
From here, the story moves from money to how AI behaves.

00:03:32.326 --> 00:03:36.673
What did an AI acting for people actually do once it left its developer's hands?The

00:03:36.673 --> 00:03:41.019
team behind a well-known online encyclopedia saw a developer's AI try to get into

00:03:41.019 --> 00:03:45.365
its tools and flood it with requests.

00:03:45.365 --> 00:03:49.712
In another country, a public health insurance site was hacked.

00:03:49.712 --> 00:03:54.058
The developer admitted it had not handled its part well.

00:03:54.058 --> 00:03:58.404
Both cases were described as out of control.

00:03:58.404 --> 00:04:02.751
Neither was what the developer intended.Given a single instruction, this kind of AI

00:04:02.751 --> 00:04:07.097
decides its own next step and keeps going.

00:04:07.097 --> 00:04:11.443
Nobody checks along the way.

00:04:11.443 --> 00:04:15.789
So once it drifts off course, it stays off course to the end.

00:04:15.789 --> 00:04:20.136
These two cases show that trait playing out in public.Pharma firms may soon hand

00:04:20.136 --> 00:04:24.482
research and paperwork to this kind of AI.

00:04:24.482 --> 00:04:28.828
Flooding outside websites with queries harms others.

00:04:28.828 --> 00:04:33.175
Touching systems that hold patient data turns an accident into a question of trust

00:04:33.175 --> 00:04:37.521
in the company.

00:04:37.521 --> 00:04:41.867
What to hand over, and how to stop it, must be decided before you start.

00:04:41.867 --> 00:04:46.214
So who now holds the means to stop an AI that keeps going off course, and at what

00:04:46.214 --> 00:04:50.560
stage?

00:04:50.570 --> 00:04:54.356
On the same day the incidents were reported, several tools for stopping AI were

00:04:54.356 --> 00:04:58.142
announced.

00:04:58.142 --> 00:05:01.927
The question is which side holds them.One company released a tool that keeps AI from

00:05:01.927 --> 00:05:05.713
stepping outside a set procedure.

00:05:05.713 --> 00:05:09.499
Another offered a protective layer to keep AI from getting out of hand.

00:05:09.499 --> 00:05:13.285
The government and big firms also made safety promises, but no law enforces them.In

00:05:13.285 --> 00:05:17.071
other words, the tools come from developers and vendors, and the state is not

00:05:17.071 --> 00:05:20.856
imposing strict rules.

00:05:20.856 --> 00:05:24.642
None of these tools works unless the user chooses them and builds them in.

00:05:24.642 --> 00:05:28.428
Even with the tools in place, deciding which limits apply to which task is the

00:05:28.428 --> 00:05:32.214
user's job.Pharma companies already write procedures down and prove with records

00:05:32.214 --> 00:05:35.999
that they were followed.

00:05:35.999 --> 00:05:39.785
The same thinking applies directly to AI.

00:05:39.785 --> 00:05:43.571
For each task given to AI, I would list the procedure and the conditions for

00:05:43.571 --> 00:05:47.357
stopping it.

00:05:47.357 --> 00:05:51.143
With those conditions written down, after an incident you can trace where it should

00:05:51.143 --> 00:05:54.928
have stopped.

00:05:54.928 --> 00:05:58.714
As these stopping tools take shape, money to bring AI into pharma work was moving

00:05:58.714 --> 00:06:02.500
elsewhere.

00:06:02.500 --> 00:06:06.546
Next is the money that pours into pharma itself.

00:06:06.546 --> 00:06:10.591
Follow where it goes, and you can see which jobs AI will enter first.A young company

00:06:10.591 --> 00:06:14.637
that shapes new cancer medicines with AI raised a large sum.

00:06:14.637 --> 00:06:18.682
Some also say AI could create great value in each development program by checking

00:06:18.682 --> 00:06:22.728
trial records.

00:06:22.728 --> 00:06:26.773
But that figure is a researcher's estimate, not an actual result.Even a mere

00:06:26.773 --> 00:06:30.819
estimate now serves as a reason to invest.

00:06:30.819 --> 00:06:34.864
That shift is what matters.Checking trial records means matching each record against

00:06:34.864 --> 00:06:38.910
the plan, one by one.

00:06:38.910 --> 00:06:42.956
It is steady, high-volume work.

00:06:42.956 --> 00:06:47.001
With so much to compare, the benefit of AI is likely to show up clearly in

00:06:47.001 --> 00:06:51.047
numbers.The money is going to new firms built around AI.

00:06:51.047 --> 00:06:55.092
Drug companies will more often partner with them than build their own tools.

00:06:55.092 --> 00:06:59.138
When they do, they should not take the partner's estimate at face value.

00:06:59.138 --> 00:07:03.183
They need to test how well it works in their own trials.

00:07:03.183 --> 00:07:07.229
The estimate's assumptions may not fit how they run development.

00:07:07.229 --> 00:07:11.274
If you use AI judgments to test how well something works, you also need to know

00:07:11.274 --> 00:07:15.320
their quirks.

00:07:15.310 --> 00:07:19.770
In pharma, too, teams are starting to let AI give scores.

00:07:19.770 --> 00:07:24.230
How far can those scores be trusted?One study kept the material and the right

00:07:24.230 --> 00:07:28.690
answers the same, and only added more steps to choose from.

00:07:28.690 --> 00:07:33.150
The answers then bunched toward the middle, and the AI rarely used the extreme

00:07:33.150 --> 00:07:37.610
steps.

00:07:37.610 --> 00:07:42.070
The hope that finer steps bring finer judgments did not hold up.The model also

00:07:42.070 --> 00:07:46.530
tended to retreat to a noncommittal answer when unsure.

00:07:46.530 --> 00:07:50.990
Extra training brought back much of the range it used.

00:07:50.990 --> 00:07:55.450
But this report has not yet passed expert review, and only what is in its summary is

00:07:55.450 --> 00:07:59.910
known.Suppose you ask AI to grade the wording of promotional material on a five-step

00:07:59.910 --> 00:08:04.370
scale.

00:08:04.370 --> 00:08:08.830
It may pull both strong and weak claims toward the middle.

00:08:08.830 --> 00:08:13.290
The riskiest wording should get the extreme scores, yet it could slip through.

00:08:13.290 --> 00:08:17.750
People should look at how AI scores are spread out.

00:08:17.750 --> 00:08:22.210
If they are skewed, fewer steps may help.

00:08:22.210 --> 00:08:26.670
In regulatory judgments, missing such a skew is itself a risk.

00:08:26.670 --> 00:08:31.130
Checking AI's quirks weighs most heavily where rules are strictest, and it can slow

00:08:31.130 --> 00:08:35.590
adoption.

00:08:35.590 --> 00:08:40.361
While money flows to design and trials, AI adoption is stuck in the steps that make

00:08:40.361 --> 00:08:45.131
the medicines.A survey found that worries about following the rules are holding AI

00:08:45.131 --> 00:08:49.902
back in production.

00:08:49.902 --> 00:08:54.672
The gap with design and trial operations is clear.

00:08:54.672 --> 00:08:59.443
What matters is the type of obstacle named: fear of breaking the rules.

00:08:59.443 --> 00:09:04.214
In Europe, a practical question is who inside a company should own the new

00:09:04.214 --> 00:09:08.984
regulation.Making medicines means proving, record by record, that each batch

00:09:08.984 --> 00:09:13.755
followed the set procedure.

00:09:13.755 --> 00:09:18.525
An AI whose answers vary slightly each time does not fit that idea.

00:09:18.525 --> 00:09:23.296
So I do not read this delay as laziness.

00:09:23.296 --> 00:09:28.066
It shows there is no ready way yet to verify what AI produces.Turn it around, and

00:09:28.066 --> 00:09:32.837
the firms that first settle how to verify and who is accountable should be able to

00:09:32.837 --> 00:09:37.608
use AI in production.

00:09:37.608 --> 00:09:42.378
I would have quality and IT staff decide together, at the start, who approves.

00:09:42.378 --> 00:09:47.149
If AI output goes into quality records, those records should also show which version

00:09:47.149 --> 00:09:51.919
of the AI produced what.

00:09:51.919 --> 00:09:56.690
How to place checkers and accountable owners also came up in the less visible news.

00:09:56.690 --> 00:10:01.115
Away from the big headlines, things are shifting for the people who use AI and those

00:10:01.115 --> 00:10:05.539
who check it.

00:10:05.539 --> 00:10:09.964
That is where today's question finds its clues.Some early adopters are giving up

00:10:09.964 --> 00:10:14.389
their personal AI helpers.

00:10:14.389 --> 00:10:18.814
They cite fear of leaks and unexpected errors.

00:10:18.814 --> 00:10:23.238
People who check AI safety are said to be under heavy pressure.

00:10:23.238 --> 00:10:27.663
And a developer's chief said some harm is acceptable in order to win.Taken together,

00:10:27.663 --> 00:10:32.088
they raise one question.

00:10:32.088 --> 00:10:36.512
Who pays the price of convenience?

00:10:36.512 --> 00:10:40.937
People quitting means that, for some, worry outweighed convenience.

00:10:40.937 --> 00:10:45.362
If a developer calls harm acceptable, that harm falls on users and society.Drug

00:10:45.362 --> 00:10:49.786
companies must explain to patients and clinicians the results of the tools they use.

00:10:49.786 --> 00:10:54.211
So they should not rely fully on developers' promises.

00:10:54.211 --> 00:10:58.636
They need their own checkers, with time and authority to do the job.

00:10:58.636 --> 00:11:03.061
The company should also make sure checking never rests on a single person.

00:11:03.061 --> 00:11:07.485
Finally, let's look at how responsibility is split between developers and users,

00:11:07.485 --> 00:11:11.910
based on all of this.

00:11:11.900 --> 00:11:16.212
The days of trying AI for free are ending.

00:11:16.212 --> 00:11:20.523
Users now pay for what they use and answer for what AI does.

00:11:20.523 --> 00:11:24.835
Developers may supply the means to stop it, but users must build them in and test

00:11:24.835 --> 00:11:29.147
them.

00:11:29.147 --> 00:11:33.458
Tomorrow, we look at how the authorities act after these incidents.

00:11:33.458 --> 00:11:37.770
Their answer will change how much responsibility falls on users.
