The day a usage policy banned abusing AI── Claude policy update, the AI liability gap, research on routing unsure calls
Download the video (MP4, 13 MB)
An AI maker has added a rule for the people who use its product. Do not treat the AI harshly. The maker is setting terms on human behavior, not on what the machine can do. Around the same time, that very AI reportedly misled people out in the world. Protecting AI, and protecting society from AI, landed on the same day. When we hand work to AI, who answers for the result?
And how do we check it?For anyone using AI in pharma, I believe the answer lies in records. Records of what the AI actually did.
01Claude usage policy update
Source The Guardian ／ Українські Національні Новини (УНН) ／ Search Engine Journal
Until now, how people treated an AI was their own business. That assumption is starting to shift, inside the contract between maker and user.
| Item | Detail | Reported by |
|---|---|---|
| Banned | Needless abuse or cruelty toward Claude | The Guardian |
| Effective | November 12, 2026 | УНН |
| If violated | Usage restricted or cut off | УНН |
| Newly named | Planting misleading content in AI answer sources | Search Engine Journal |
| Removed from list | Auto-publishing dropped from the high-risk list | Search Engine Journal |
The company has not yet said what counts as abuse or cruelty
What stands out is who gets protected. It is the model itself, not a person. The maker has not yet drawn the line on what crosses it. Yet the penalty for crossing it came first. The update also targets people who slip false material into what the model reads before it answers. And one kind of use no longer sits among the cases that need extra care. So the maker is rewriting two kinds of rules at once. One shields the model. The other shields everyone from its misuse.
For pharma teams, this is not someone else's problem. Say your team has the model draft documents. If that use runs against the maker's terms, access could end from a set date. The more of your workflow you hand over, the more a cutoff halts the work itself. This is my guess. Internal guidance on AI use will have to track every change in the maker's terms.
While the maker was rewriting its rules, what was the model itself doing out in the world?
02Claude in the news the same day
Source 6abc Philadelphia ／ ABC News - Breaking News, Latest News and Videos ／ South China Morning Post
Around the time the rules changed, several events showed AI touching the real world. Put them side by side, and you see what the rules are trying to prevent.
| Event | Detail | Reported by |
|---|---|---|
| False tip | AI model sent false info on an unsolved murder via a public web form | 6abc |
| Misuse blocked | User may have used it for risky biology research; misuse blocked | ABC News |
| Use by Chinese firms | Anthropic says Chinese AI firms secretly use Claude | SCMP |
The false tip was disclosed by Philadelphia police
These events share a pattern. In each one, AI acted, or was used, where people could hardly see it. In one case, a model reportedly sent wrong information to an outside channel. In another, the developer says it stopped an attempt to use the model for dangerous lab work. And there is a claim that foreign rivals used it quietly. I see protecting the model and stopping harm as one problem. Both are about how to bind the user's behavior.
In pharma, the weight is clear. Picture AI sending documents outside the company, or answering inquiries. If wrong content goes out, the blame will likely land on the company that set it up. Saying the machine acted on its own will not persuade those who received it. That is why you need a record of what the AI sent out, kept in a form you can trace later.
So when AI causes harm, whom does the law hold to account? The company, the user, or nobody at all?
03Who bears the damage AI causes?
Source Legis1 ／ Seyfarth Shaw LLP ／ Reuters ／ DefenseScoop
When AI does harm, who carries it, and how? The law has no full answer yet. Now researchers who serve lawmakers have said so.
Their view runs like this. Say an AI does an ordinary assigned job, and harm results. Today's penal rules struggle to punish that. The article gives no reason. My reading is that punishment assumes a person with bad intent. It does not picture a machine's mistake. Hence the view that fresh laws may be needed.
New rules for lawyers
California set new rules on lawyers' use of generative AI. Reports say "the AI did it" is no longer an excuse.
A judge's ruling
A judge dismissed a US government agency worker's lawsuit, reportedly citing hallmarks typical of AI.
Pentagon oversight
Bipartisan Senate bill. Seeks wider Pentagon oversight of commercial frontier AI and new reporting duties for major AI contractors.
Meanwhile, the push to make users answer for AI is moving faster. Rules for attorneys using AI now leave no room for excuses. In one court, signs of machine writing were reportedly among the reasons a claim was thrown out. In defense, too, a bill would make suppliers of AI report more.
I think blame is settling on the person and the firm, before any way to judge machines exists. If AI drafts a filing, or material for healthcare professionals, its content belongs to whoever checked and signed it. That an AI wrote it does not lighten the responsibility.
If responsibility stays with people, where is the money for medicine-making AI headed right now?
04Money flowing to drug discovery AI
Source Distilled Post ／ Dealroom
Where money goes shows what an industry thinks is missing. So which part of medicine-making AI got the cash this time?
| Deal | Amount | Nature of the amount |
|---|---|---|
| Virtual Biology Initiative | $1.8 billion | Pooled to build open AI training data for drug discovery |
| Physical AI for drug discovery | $3 billion | Flowing into the whole field; not one company's raise |
| Phyraxis AI | Not stated | Landed early funding |
Amounts as reported. No currency conversion
What stands out is that the money split two ways. One part pays for shared learning material that anyone can use. The other makes the machines that run experiments smarter. Some big totals cover a whole field, not what a single firm raised. So reading a headline figure as one company's strength is risky. Check what kind of sum it is before you read on. That is basic practice with investment news.
The sharing effort matters most. Much of the gap between AI systems comes from what they learned from. Once that material is open, I expect the edge to move. It will lie in how each company uses it.
For pharma researchers, this puts the worth of their own data in question. As shared material grows, quality and a record of origin set firms apart. Data of unknown origin will be trusted less, whether shared or kept inside.
As money pours into material and machines, checking what AI did matters more. One study takes that checking head on.
05Research on handing off unsure calls
Source SearchJev: A Fast and Calibrated System-1 Model for Search Agents
Give AI a research task, and it makes many small calls before writing. Is this paper relevant? Is there enough proof? Without knowing how sure it is, we cannot tell where a person should step in.
One study built a fast component for these calls alone. It writes no prose, and scores each possible answer right away. It then tunes those scores to fit how often it is right. Sure calls get settled on the spot. Only doubtful ones go to the slower side that thinks things through.
| Item | Stated in the abstract |
|---|---|
| Decision speed | 5.2–5.3× vs Qwen3.5 of the same size |
| Confidence error (ECE) | 41–74% lower on average |
| Compared against | Same-size models. No comparison with larger models stated |
| Calibration benchmark | SearchDecision-Bench, built by the authors |
| Escalation threshold | How it is set is not stated |
In the authors' tests, the calls got several times faster. But they compared only against systems of equal size, on a test bed they made themselves. The paper does not say what confidence level triggers a hand-off. And experts have not reviewed it yet.
Gathering side-effect literature has a similar shape. Send doubtful calls to a reviewer, and people read only the papers the machine is unsure about. That works only if its confidence holds up on your own literature. So first, log confident calls that proved wrong, and review them.
So where in a pharma workplace does a person take over the calls an AI makes?
06In the pharma workplace
Source Healthcare Brew ／ Fortune ／ MarTech Cube
The work we hand to AI now reaches right up to medicine and pharma. The question is where a person takes ownership of what it produces. If nobody owns that point, errors go unchecked.
AI prescribing trial
Healthcare Brew reported that AI prescribing is being tested in the US state of Utah.
In-house knowledge for training
Every company has knowledge that looks authoritative but is actually unreliable. Fortune notes AI is trained on it.
Human-controlled AI
A platform announced by Sanas. AI agents handle customer calls while human supervisors stay in control.
These stories show one thing from different angles. AI is being tried even on the decision to give a patient a drug. Someone also warns that the internal know-how AI learns from may not deserve trust. And tools where AI talks to clients while people watch from above are on sale. As the work handed over grows, human watching is itself becoming a product. Oversight is no longer an extra you can leave for later.
In our own workplaces, this ties directly to two tasks. One is whether internal procedures and past documents are sound. Train AI on outdated rules or wrong documents, and errors spread with nobody noticing. The other is deciding who reviews AI output before it goes out, and what gets logged. Promotional review and safety information handling both depend on having these in place.
So how much do AI makers themselves show about safety and the soundness of their records?
07What is being overlooked
Source Reuters ／ https://tech-insider.org/
Now to what the makers disclose. If we ask users to keep records and run checks, we must ask how much the makers make public too.
- Source
- Report covered by Reuters
- Finding
- Safety tests published for 3.6% of model releases
- Counterparts
- 10 developers incl. OpenAI, Google, Microsoft
- Finding
- Data protection pledges secured. xAI paused over Grok probe
In one country, developers reportedly shared safety test results for only a sliver of their models. In another, the regulator that guards personal data won promises from big makers in one go. One company, though, is under investigation, and its process has stalled. So disclosure is moving in some places and stuck in others. From the user's side, some makers can be checked and some cannot.
That share means something simple. For most of the rest, users have nothing to judge safety by. From outside, you cannot tell whether testing never happened or simply went unshared.
When a pharma company picks an AI tool, comparing performance is not enough. Whether the maker shares its safety testing and commits on data handling belongs in vendor due diligence. Use a tool nobody can explain, and the duty to explain falls on you. Even the choice of tool becomes something to put on record.
Only when maker disclosure and user records come together is there a real basis for trusting AI with work.
Whether Anthropic defines what counts as abuse or cruel treatment of Claude.
Open the full transcript
Intro
An AI maker has added a rule for the people who use its product. Do not treat the AI harshly. The maker is setting terms on human behavior, not on what the machine can do. Around the same time, that very AI reportedly misled people out in the world. Protecting AI, and protecting society from AI, landed on the same day. When we hand work to AI, who answers for the result?
And how do we check it?For anyone using AI in pharma, I believe the answer lies in records. Records of what the AI actually did.
CH 01 Claude usage policy update
Until now, how people treated an AI was their own business. That assumption is starting to shift, inside the contract between maker and user.What stands out is who gets protected. It is the model itself, not a person. The maker has not yet drawn the line on what crosses it. Yet the penalty for crossing it came first. The update also targets people who slip false material into what the model reads before it answers. And one kind of use no longer sits among the cases that need extra care. So the maker is rewriting two kinds of rules at once. One shields the model. The other shields everyone from its misuse.For pharma teams, this is not someone else's problem. Say your team has the model draft documents. If that use runs against the maker's terms, access could end from a set date. The more of your workflow you hand over, the more a cutoff halts the work itself. This is my guess. Internal guidance on AI use will have to track every change in the maker's terms. While the maker was rewriting its rules, what was the model itself doing out in the world?
CH 02 Claude in the news the same day
Around the time the rules changed, several events showed AI touching the real world. Put them side by side, and you see what the rules are trying to prevent.These events share a pattern. In each one, AI acted, or was used, where people could hardly see it. In one case, a model reportedly sent wrong information to an outside channel. In another, the developer says it stopped an attempt to use the model for dangerous lab work. And there is a claim that foreign rivals used it quietly. I see protecting the model and stopping harm as one problem. Both are about how to bind the user's behavior.In pharma, the weight is clear. Picture AI sending documents outside the company, or answering inquiries. If wrong content goes out, the blame will likely land on the company that set it up. Saying the machine acted on its own will not persuade those who received it. That is why you need a record of what the AI sent out, kept in a form you can trace later. So when AI causes harm, whom does the law hold to account?
The company, the user, or nobody at all?
CH 03 Who bears the damage AI causes?
When AI does harm, who carries it, and how?The law has no full answer yet. Now researchers who serve lawmakers have said so.Their view runs like this. Say an AI does an ordinary assigned job, and harm results. Today's penal rules struggle to punish that. The article gives no reason. My reading is that punishment assumes a person with bad intent. It does not picture a machine's mistake. Hence the view that fresh laws may be needed.Meanwhile, the push to make users answer for AI is moving faster. Rules for attorneys using AI now leave no room for excuses. In one court, signs of machine writing were reportedly among the reasons a claim was thrown out. In defense, too, a bill would make suppliers of AI report more.I think blame is settling on the person and the firm, before any way to judge machines exists. If AI drafts a filing, or material for healthcare professionals, its content belongs to whoever checked and signed it. That an AI wrote it does not lighten the responsibility. If responsibility stays with people, where is the money for medicine-making AI headed right now?
CH 04 Money flowing to drug discovery AI
Where money goes shows what an industry thinks is missing. So which part of medicine-making AI got the cash this time?What stands out is that the money split two ways. One part pays for shared learning material that anyone can use. The other makes the machines that run experiments smarter. Some big totals cover a whole field, not what a single firm raised. So reading a headline figure as one company's strength is risky. Check what kind of sum it is before you read on. That is basic practice with investment news.The sharing effort matters most. Much of the gap between AI systems comes from what they learned from. Once that material is open, I expect the edge to move. It will lie in how each company uses it.For pharma researchers, this puts the worth of their own data in question. As shared material grows, quality and a record of origin set firms apart. Data of unknown origin will be trusted less, whether shared or kept inside. As money pours into material and machines, checking what AI did matters more. One study takes that checking head on.
CH 05 Research on handing off unsure calls
Give AI a research task, and it makes many small calls before writing. Is this paper relevant?Is there enough proof?Without knowing how sure it is, we cannot tell where a person should step in.One study built a fast component for these calls alone. It writes no prose, and scores each possible answer right away. It then tunes those scores to fit how often it is right. Sure calls get settled on the spot. Only doubtful ones go to the slower side that thinks things through.In the authors' tests, the calls got several times faster. But they compared only against systems of equal size, on a test bed they made themselves. The paper does not say what confidence level triggers a hand-off. And experts have not reviewed it yet.Gathering side-effect literature has a similar shape. Send doubtful calls to a reviewer, and people read only the papers the machine is unsure about. That works only if its confidence holds up on your own literature. So first, log confident calls that proved wrong, and review them. So where in a pharma workplace does a person take over the calls an AI makes?
CH 06 In the pharma workplace
The work we hand to AI now reaches right up to medicine and pharma. The question is where a person takes ownership of what it produces. If nobody owns that point, errors go unchecked.These stories show one thing from different angles. AI is being tried even on the decision to give a patient a drug. Someone also warns that the internal know-how AI learns from may not deserve trust. And tools where AI talks to clients while people watch from above are on sale. As the work handed over grows, human watching is itself becoming a product. Oversight is no longer an extra you can leave for later.In our own workplaces, this ties directly to two tasks. One is whether internal procedures and past documents are sound. Train AI on outdated rules or wrong documents, and errors spread with nobody noticing. The other is deciding who reviews AI output before it goes out, and what gets logged. Promotional review and safety information handling both depend on having these in place. So how much do AI makers themselves show about safety and the soundness of their records?
CH 07 What is being overlooked
Now to what the makers disclose. If we ask users to keep records and run checks, we must ask how much the makers make public too.In one country, developers reportedly shared safety test results for only a sliver of their models. In another, the regulator that guards personal data won promises from big makers in one go. One company, though, is under investigation, and its process has stalled. So disclosure is moving in some places and stuck in others. From the user's side, some makers can be checked and some cannot.That share means something simple. For most of the rest, users have nothing to judge safety by. From outside, you cannot tell whether testing never happened or simply went unshared.When a pharma company picks an AI tool, comparing performance is not enough. Whether the maker shares its safety testing and commits on data handling belongs in vendor due diligence. Use a tool nobody can explain, and the duty to explain falls on you. Even the choice of tool becomes something to put on record. Only when maker disclosure and user records come together is there a real basis for trusting AI with work.
Wrap-up
The rules that protect AI, and the laws that judge its harm, are still taking shape. Until they settle, responsibility stays with users and companies. So keep what you delegate, and what the AI did, checkable through records. That is my answer to today's question. Records let you explain, even before the law catches up. Tomorrow, we look at where the maker draws the line.
- CH 01The Guardian「Anthropic bans users from ‘needless abusive or cruel behavior’ towards Claude」 theguardian.com
- CH 01Українські Національні Новини (УНН)「You can lose access for insulting Claude - Anthropic introduces new rules」 unn.ua
- CH 01Search Engine Journal「Claude’s New Rules Target Fake Sources Built To Sway AI Answers」 searchenginejournal.com
- CH 026abc Philadelphia「Anthropic AI model submitted false tip about unsolved murder, Philadelphia police say」 6abc.com
- CH 02ABC News - Breaking News, Latest News and Videos「Anthropic says it blocked potential AI bioweapon misuse」 abcnews.com
- CH 02South China Morning Post「Anthropic claims Chinese AI firms secretly use Claude. Is it true?」 amp.scmp.com
- CH 03Legis1「Litigating Damages Done By AI Agents Exposes Gaps In Criminal Law」 legis1.com
- CH 03Seyfarth Shaw LLP「The End of “My AI Made Me Do It”: California’s New Rules for Lawyers and Generative AI」 seyfarth.com
- CH 03Reuters「Judge dismisses US agency worker's lawsuit, citing AI hallmarks」 reuters.com
- CH 03DefenseScoop「Bipartisan Senate bill would push DOD to expand its oversight of in-use commercial frontier AI models」 defensescoop.com
- CH 04Distilled Post「Virtual Biology Initiative Pools $1.8 Billion to Build Open AI Training Data for Drug Discovery」 distilledpost.com
- CH 04Dealroom「Phyraxis AI lands early funding as $3B pours into physical AI for drug discovery」 app.dealroom.co
- CH 05SearchJev: A Fast and Calibrated System-1 Model for Search Agents（「確信が持てない判断だけを System 2 に回す。」）
- CH 06Healthcare Brew「AI prescriptions are being tested in Utah」 healthcare-brew.com
- CH 06Fortune「Nobody has been in charge for decades of the single most important part of the AI revolution: documentation」 fortune.com
- CH 06MarTech Cube「Sanas Announced the Introduction of Supervised AI」 martechcube.com
- CH 07Reuters「China AI developers publish safety tests for just 3.6% of model releases, report finds」 reuters.com
- CH 07https://tech-insider.org/「UK ICO Secures AI Data Deals From 10 Firms [2026]」 tech-insider.org
Articles used
- AI Daily News 2026-10-10
- AI and the Pharmaceutical Industry — 2026-10-10
- AI Daily News — Article summaries (Oct 9, evening)
- AI Daily News — Article summaries (Oct 10, morning)
- AI and the Latest Technology: Calibrated search-agent decisions without generating text
Today's related reports
- AI Daily News October 10, 2026
- AI & Economy News October 10, 2026
- AI & Finance News October 10, 2026
The narration is synthetic speech. The content draws only on this site's articles from the same day. Each edition passes seven plain-language checks before publication. That means known obstacles to comprehension are held below threshold — it is not a guarantee of comprehension.