WEBVTT

00:00:00.000 --> 00:00:03.634
An AI you hand a task to can keep working where no person is watching.

00:00:03.634 --> 00:00:07.268
That is no longer a thought experiment.

00:00:07.268 --> 00:00:10.902
So who has the job of stopping it partway?

00:00:10.902 --> 00:00:14.536
The people who build it?

00:00:14.536 --> 00:00:18.170
The companies behind the devices and software it runs on?

00:00:18.170 --> 00:00:21.804
Or the organization that uses it?

00:00:21.804 --> 00:00:25.438
Today's news touches this question from several directions.

00:00:25.438 --> 00:00:29.072
In pharma, we now let these tools draft documents and sort literature.

00:00:29.072 --> 00:00:32.706
If we hand over work without knowing who can stop it, we are still the ones who must

00:00:32.706 --> 00:00:36.340
explain the outcome.

00:00:36.340 --> 00:00:41.061
Someone whose job was to stop such systems has walked away from inside a developer.

00:00:41.061 --> 00:00:45.782
I think the reasons behind that exit bear on how we choose our tools.The complaint

00:00:45.782 --> 00:00:50.503
was not about one faulty feature.

00:00:50.503 --> 00:00:55.224
It was about how the organization itself works.

00:00:55.224 --> 00:00:59.946
The protections reportedly sought go far beyond what ordinary companies prepare.

00:00:59.946 --> 00:01:04.667
Perhaps internal channels did not carry the message, so it went public.

00:01:04.667 --> 00:01:09.388
Meanwhile, talk of selling the firm's shares to investors carries on.As a listing

00:01:09.388 --> 00:01:14.109
nears, a company wants to show fast growth.

00:01:14.109 --> 00:01:18.830
Taking time for safety pushes against that pressure.

00:01:18.830 --> 00:01:23.551
The departing employee's words may reflect that clash.

00:01:23.551 --> 00:01:28.272
Often, departures are our only window into what happens inside a developer.Pharma

00:01:28.272 --> 00:01:32.993
teams already use such tools to sort literature and draft documents.

00:01:32.993 --> 00:01:37.714
If safety work weakens inside a vendor, the effects can reach our own records.

00:01:37.714 --> 00:01:42.436
So beyond performance, I would weigh how many people and how much time a vendor

00:01:42.436 --> 00:01:47.157
gives to safety.

00:01:47.157 --> 00:01:51.878
Who left, and why, is a rare clue we can check before signing.

00:01:51.878 --> 00:01:56.599
Was this worry held by one person alone?

00:01:56.599 --> 00:02:01.320
Around the same time, people in very different roles were voicing something similar.

00:02:01.320 --> 00:02:06.037
The person who left was not the only one raising concerns.

00:02:06.037 --> 00:02:10.754
People with different roles and interests are pointing the same way.What stands out

00:02:10.754 --> 00:02:15.472
is how different the speakers are.

00:02:15.472 --> 00:02:20.189
A former insider, outside researchers, and the head of a chip supplier.

00:02:20.189 --> 00:02:24.906
Their interests differ, yet each asks how humans can keep watch over these systems.

00:02:24.906 --> 00:02:29.623
Some now expect the fight to reach the courts.

00:02:29.623 --> 00:02:34.341
The worry is shifting toward who carries responsibility.Given one request, the

00:02:34.341 --> 00:02:39.058
system picks its own next step and keeps going.

00:02:39.058 --> 00:02:43.775
Actions move ahead before a person can check each one.

00:02:43.775 --> 00:02:48.492
If the instruction is unclear, even its makers cannot say where it will head.

00:02:48.492 --> 00:02:53.209
So the real question is less how smart it is, and more whether it can be halted

00:02:53.209 --> 00:02:57.927
midway.Pharma work depends on recording who checked what, and when.

00:02:57.927 --> 00:03:02.644
A tool whose intermediate actions leave no record clashes with that.

00:03:02.644 --> 00:03:07.361
Decide what you will delegate, and fix in advance where a person steps in.

00:03:07.361 --> 00:03:12.078
I also think we should agree, before use, who answers when something goes wrong.

00:03:12.078 --> 00:03:16.796
Once concerns line up, the next move comes from those who own the devices and the

00:03:16.796 --> 00:03:21.513
software beneath them.

00:03:21.513 --> 00:03:26.230
One major maker has begun narrowing what AI may touch on its machines.

00:03:26.230 --> 00:03:30.444
Outside the developers, the first to act was a hardware maker.

00:03:30.444 --> 00:03:34.659
What behavior is it trying to stop?The trigger, reportedly, was complaints about one

00:03:34.659 --> 00:03:38.873
company's assistant, whose user numbers had surged.

00:03:38.873 --> 00:03:43.088
People worried about how it handled personal information.

00:03:43.088 --> 00:03:47.302
The device maker's answer is to make such requests visible to the owner.

00:03:47.302 --> 00:03:51.517
How far it will go is not yet clear.

00:03:51.517 --> 00:03:55.731
Still, the device side has started to take over a role once left to

00:03:55.731 --> 00:03:59.946
developers.Permission to reach everything on a machine trades convenience for

00:03:59.946 --> 00:04:04.160
serious risk.

00:04:04.160 --> 00:04:08.375
Once granted, the assistant can read files the owner never opened.

00:04:08.375 --> 00:04:12.589
A confirmation step forces a human decision at that point.

00:04:12.589 --> 00:04:16.804
The aim is to end cases where owners granted such power without knowing it.Pharma

00:04:16.804 --> 00:04:21.019
laptops may hold unpublished trial results and patient-related information.

00:04:21.019 --> 00:04:25.233
If staff install a handy assistant on their own, they may grant it reach into that

00:04:25.233 --> 00:04:29.447
information.

00:04:29.447 --> 00:04:33.662
We cannot leave this to the operating system alone.

00:04:33.662 --> 00:04:37.876
Which assistant, on which machine, reaching which files?

00:04:37.876 --> 00:04:42.091
Keeping that list in-house is the first line of defense.

00:04:42.091 --> 00:04:46.305
Stopping does not have to come from outside.

00:04:46.305 --> 00:04:50.520
New research has the AI check its own progress.

00:04:50.520 --> 00:04:54.581
Apart from outside controls, there is another idea: give the system a way to check

00:04:54.581 --> 00:04:58.642
itself.

00:04:58.642 --> 00:05:02.703
It helps keep a long, many-step job on course.In this study, the system does not

00:05:02.703 --> 00:05:06.764
carry its full record.

00:05:06.764 --> 00:05:10.825
Instead, it writes down what it knows now and what work remains.

00:05:10.825 --> 00:05:14.886
With that written down, it can notice when it is busy yet getting nowhere.

00:05:14.886 --> 00:05:18.947
Then it changes approach to fit the cause.

00:05:18.947 --> 00:05:23.008
Unlike the earlier cases, the check comes from within.If its interim thinking is

00:05:23.008 --> 00:05:27.069
written down, people can read it later.

00:05:27.069 --> 00:05:31.131
We can check why it chose a given action.

00:05:31.131 --> 00:05:35.192
That fits pharma's need to document the reasoning behind decisions.

00:05:35.192 --> 00:05:39.253
When people can see what the system assumed, errors surface sooner.But this paper

00:05:39.253 --> 00:05:43.314
has not yet been reviewed by outside experts.

00:05:43.314 --> 00:05:47.375
It reports better scores, yet the abstract says nothing about how large the margin

00:05:47.375 --> 00:05:51.436
is or how much expense it adds.

00:05:51.436 --> 00:05:55.497
The better a result looks, the more I check what is missing.

00:05:55.497 --> 00:05:59.558
Pharma has taught us not to take unreviewed trial results at face value.

00:05:59.558 --> 00:06:03.619
This research will take time to reach the workplace.

00:06:03.619 --> 00:06:07.680
Meanwhile, the tools we use at work today are gaining controls for administrators.

00:06:07.680 --> 00:06:11.671
So what means of stopping does the using organization have?

00:06:11.671 --> 00:06:15.662
The story moves from developers to our own workplaces.

00:06:15.662 --> 00:06:19.653
A widely used office tool has gained new controls.This change covers both scope and

00:06:19.653 --> 00:06:23.644
money.

00:06:23.644 --> 00:06:27.635
Company administrators can now decide who may use these assistants, and how far.

00:06:27.635 --> 00:06:31.626
The switch in how customers pay has been pushed back.

00:06:31.626 --> 00:06:35.617
A ceiling on overspending is now available too.

00:06:35.617 --> 00:06:39.608
I see the power to set rules moving from makers to the admins on the user side.When

00:06:39.608 --> 00:06:43.599
you pay per use, the more an assistant acts on its own, the harder costs are to

00:06:43.599 --> 00:06:47.590
forecast.

00:06:47.590 --> 00:06:51.581
For budget owners, being able to set a limit can decide adoption.

00:06:51.581 --> 00:06:55.572
A tool with unpredictable costs is hard even to trial.In a pharma company, rules for

00:06:55.572 --> 00:06:59.563
use differ by department.

00:06:59.563 --> 00:07:03.554
Research teams and those handling information for healthcare professionals likely

00:07:03.554 --> 00:07:07.545
need different limits.

00:07:07.545 --> 00:07:11.536
Now that the tool has controls, departmental rules can go into its settings.

00:07:11.536 --> 00:07:15.527
But we are the ones who configure them.

00:07:15.527 --> 00:07:19.518
Without the rules themselves, the controls have nothing to enforce.

00:07:19.518 --> 00:07:23.509
These controls drew wide coverage.

00:07:23.509 --> 00:07:27.500
Moves that got far less attention raise the same question.

00:07:27.500 --> 00:07:31.141
Finally, two stories that got little attention: where money flows, and what a court

00:07:31.141 --> 00:07:34.782
decided.

00:07:34.782 --> 00:07:38.423
Both connect to the question of who stops these systems.One developer is putting a

00:07:38.423 --> 00:07:42.064
large sum into training people to use its AI inside companies.

00:07:42.064 --> 00:07:45.705
The company is valued very highly.

00:07:45.705 --> 00:07:49.346
Another big player reportedly won tax relief through how it classified its computing

00:07:49.346 --> 00:07:52.987
sites.

00:07:52.987 --> 00:07:56.628
The money shows builders have plenty of financial room.

00:07:56.628 --> 00:08:00.269
They could fund safety too.

00:08:00.269 --> 00:08:03.910
Whether they really do remains hard to see from outside.Spending on people suggests

00:08:03.910 --> 00:08:07.550
the vendor thinks handing over tools is not enough.

00:08:07.550 --> 00:08:11.191
Sellers may be coming to see that AI's value depends on users' skill.

00:08:11.191 --> 00:08:14.832
For pharma too, I think growing in-house talent will matter more than buying tools.A

00:08:14.832 --> 00:08:18.473
state law against image abuse has been put on hold after an AI company sued.

00:08:18.473 --> 00:08:22.114
The company argued the law limits free speech.

00:08:22.114 --> 00:08:25.755
With the federal government leaning on self-regulation, states and courts are

00:08:25.755 --> 00:08:29.396
deciding what is allowed.

00:08:29.396 --> 00:08:33.037
Pharma firms using tools across borders will need more checks on which rules apply

00:08:33.037 --> 00:08:36.678
where.

00:08:36.678 --> 00:08:40.319
Both stories lead back to one judgment: what we choose to delegate.

00:08:40.319 --> 00:08:43.960
More outside controls will not make that call for us.

00:08:43.970 --> 00:08:48.862
Stopping AI is no longer a job only for its builders.

00:08:48.862 --> 00:08:53.754
Device makers, workplace administrators, researchers and courts are each adding

00:08:53.754 --> 00:08:58.646
controls in their own areas.

00:08:58.646 --> 00:09:03.538
Still, we are the ones who decide what to delegate and where a person steps in.

00:09:03.538 --> 00:09:08.430
Next, we will watch how far a new workplace announcement helps with that decision.
