Outside eyes found the hole in AI oversight── Bans, money, electricity: what pharma should check before trusting AI
Download the video (MP4, 8 MB)
Here is a surprise. The hole in one AI watchdog was not found by the company that built it. By the end, you will know what to check before you hand work to AI. Is it enough for a maker to watch its own system?
On the same day, tools that take over our work only multiplied. A developer told by a government to stay out reportedly raised a huge sum within days. The day we trust AI with patient records and development data is coming close to pharma, too. Will the money and electricity behind it last?
The clues all point to one place.
01Outside group flags monitoring flaws
Source Reuters ／ Mashable ／ StateScoop
A safety net we relied on was found to have a hole, and from outside. For anyone handing work to AI, that shakes the ground.
What caught my eye is who found it. It was not the builder. When you test your own system, blind spots usually stay blind. An outside reviewer brought the weak point into view. A maker's own word is not enough to call its AI safe.
| Actor | Move | Reported by |
|---|---|---|
| METR | Flagged flaws in OpenAI's internal monitoring system | Reuters, CNBC |
| Sam Altman | Stepped down as head of OpenAI's safety committee | Mashable |
| OpenAI | Urged Governor Newsom to strengthen California's AI safety law | StateScoop |
That same day, the people overseeing safety changed. And the company itself asked for tougher state rules. This is only my reading. But it may sense that self-policing is not enough.
An outside reviewer has no stake in sales. It can name flaws that hurt revenue. In pharma, having someone else check your work is standard practice. AI now needs the same.
So when choosing AI, a performance pitch is not enough. Ask who tested it from outside, and whether you can see the results. Write regular reporting into the contract.
Meanwhile, while oversight wobbles, how fast are the tools that take over our work spreading?
02Agent products line up
Source Google Cloud ／ Google for Developers Blog ／ The New York Times ／ Search Engine Journal
On the very day oversight was questioned, tools that act for us only grew in number.
Claude Duet
Anthropic's AI agent. Reported by TechCrunch, Mashable and Barron's, framed as a rival to Meta's Muse
Google's context enrichment
Google Cloud announced Knowledge Catalog, a self-learning context enrichment feature for AI agents
Gemini development platform
How to build governed, self-evolving AI agents on Gemini Agent Platform; says failures recover in seconds
Every tool here is pitched as running without a human at the wheel. It learns, repairs and gathers what it needs by itself. The more convenient it gets, the fewer chances people have to step in. So the oversight problem will grow, not shrink. Checking is not keeping pace with delegation.
| Person | Statement / move |
|---|---|
| Sundar Pichai (Google) | Search will be absorbed into an agentic layer |
| Sundar Pichai (Google) | AI agents track Google's operations 24 hours a day |
| Walmart, Amazon | Expanding AI agents' role in shopping and operations |
Big tech leaders expect the doorway to information itself to become a delegated assistant. In buying and daily tasks, the screens people operate may shrink. Choices then happen inside the machine, and the reasons become hard to see.
Such a system picks its own path. So even a right answer needs a separate check on the route. In pharma, the trail matters as much as the conclusion.
Soon these tools will draft literature searches and replies to inquiries. Staff will need rules on what gets logged, more than a user manual. Decide who signs off before you start.
But rules on which tools we may use are not only internal. Sometimes a government chooses first.
03Funding right after a usage ban
Source TechCrunch
Sometimes a government decides first which AI may be used. What happened to the developer it shut out?
The order covered AI used inside federal agencies. It told them to drop one developer's technology for security reasons. Yet within days, that developer reportedly pulled in a vast sum. In this story, exclusion did not slow it down.
One government's order binds only what its own offices use. It cannot reach where a developer raises money, or who uses it worldwide.
The sum is said to rank among the largest in its home country. To me, faith in its momentum outweighed the order. Investors seem to treat the ban as a local hurdle.
Drug makers use AI across borders too. A tool allowed at one site may be barred at another. More and more, local rules will decide which AI you may use, before performance does. Map which orders could touch your tools before rollout.
If money flows even to a developer that was shut out, investment in AI remains strong. Next, the money closest to pharma.
04Drug discovery AI: what the sums mean
Source BioPharma APAC ／ Pulse 2.0 ／ Fortune ／ Yahoo Finance
Some large sums near pharma were reported too. Whether to take them at face value is another matter.
| Deal | Reported amount | Nature of the amount |
|---|---|---|
| Oncology collaboration: AI drug discoverer and drug maker | $888 Million | Headline does not show if confirmed or a total cap |
| NIH grant for an AI metabolomics drug discovery model | $2.5 Million | Public grant; small in scale |
| Legal AI firm Legora's additional raise | $550 million | Amount raised |
| Meta's fundraising for AI | $145 billion | Raised without giving up voting control |
Nature of amounts per this site's "AI and the Pharmaceutical Industry" and "AI Daily News"
At a glance, these sums look alike. They are not. Money already paid and a ceiling paid only if conditions are met mean very different things. A public grant is money that will arrive. Rank deals by size alone, and you misjudge their weight.
Drug discovery partnerships often pay in steps, each time development reaches a preset point. If development stops, the rest never comes. So the big total here is best read as a ceiling, not cash in hand.
Today's news was only that the first step was passed. Passing it guarantees nothing about the next. We need the habit of separating reported sums from money that moved.
Pharma staff often explain partnership news inside their company. Quote only the headline sum, and expectations swell. Add one line on what kind of money it is. That small habit keeps colleagues from misjudging.
So what is moving around money and AI right next to the daily work of pharma teams?
05On the pharma front line
Source The Verge ／ Bloomberg Law News ／ Law.com
Drug makers are reportedly teaming up with firms that own vast computing power, rather than building AI in house.
| Pharma's partner | Area |
|---|---|
| Oracle | Trial optimization and patient finding |
| NVIDIA, Microsoft | Support for patients with rare tumors |
Drug company names omitted. From this site's "AI and the Pharmaceutical Industry", Oct 23 issue
The partners' work reaches beyond the lab, into running human trials and helping patients. AI is entering the base that runs development itself. The more of that base you hand over, the more the partner's safety and oversight become your own quality. That is why you check their oversight first.
Diagnostic AI claim
Microsoft AI CEO Mustafa Suleyman claimed the company's AI tool beats doctors at diagnosing disease
Training data lawsuit
Independent music publishers sued Anthropic, alleging lyrics were copied without permission to train Claude
Insurance and AI governance
Documenting AI governance and risk management helps negotiate coverage and exclusions at commercial insurance renewal
The same day brought a bold claim that AI outdiagnoses physicians. A developer was sued over training material used without consent. And written AI rules may help when policies come up for review. Different stories, one question: can you explain yourself?
Pharma has always earned trust through records and explanation. If you cannot show what an AI learned from and how it is managed, the user shares the blame. Bold claims need their evidence checked. Which patients, compared with whom? It is the same question we ask of a drug.
Yet all of this assumes that the electricity and cash behind the computing keep flowing.
06Power and money as limits
Source Crusoe ／ FOX 5 New York ／ CNBC ／ Reuters
Finally, let us look at the ground under AI's endless growth.
| Actor | Move | Reason / remark |
|---|---|---|
| New York Governor Hochul | Defends pause on AI data centers | Power shortage concerns |
| Crusoe, NBSF | Power facility for an AI factory | $700 million |
| Nvidia CEO Jensen Huang | Neocloud demand is "off the charts" | Yet capital is scarce |
| Tesla | Cut capital spending | Shares fell |
One state is defending a halt to new computing sites, citing fears of running short of electricity. Others are building their own supply. Electricity is becoming one of the first things AI growth runs out of. Demand is strong, yet the cash behind it looks thin. A company that cut big investments saw its stock drop.
These new compute renters spend heavily just to buy machines. If cash gets harder to raise, their capacity grows slowly. For a renter, today's terms may not hold tomorrow. Plan for price rises, or a provider pulling out.
If drug makers move research onto AI, estimate how long that base will last. A partner's electricity and cash troubles become weak spots in your own plans. I would prepare an exit route from day one. Keep data in a format another provider can run.
Oversight, exclusion, cash and electricity. Every story today led back to one point: what to check before you hand over the work.
How will OpenAI respond to METR's findings? Watch its response to the internal monitoring flaws and where changes to California's AI safety law go.
Open the full transcript
Intro
Here is a surprise. The hole in one AI watchdog was not found by the company that built it. By the end, you will know what to check before you hand work to AI. Is it enough for a maker to watch its own system?
On the same day, tools that take over our work only multiplied. A developer told by a government to stay out reportedly raised a huge sum within days. The day we trust AI with patient records and development data is coming close to pharma, too. Will the money and electricity behind it last?
The clues all point to one place.
CH 01 Outside group flags monitoring flaws
A safety net we relied on was found to have a hole, and from outside. For anyone handing work to AI, that shakes the ground.What caught my eye is who found it. It was not the builder. When you test your own system, blind spots usually stay blind. An outside reviewer brought the weak point into view. A maker's own word is not enough to call its AI safe.That same day, the people overseeing safety changed. And the company itself asked for tougher state rules. This is only my reading. But it may sense that self-policing is not enough.An outside reviewer has no stake in sales. It can name flaws that hurt revenue. In pharma, having someone else check your work is standard practice. AI now needs the same.So when choosing AI, a performance pitch is not enough. Ask who tested it from outside, and whether you can see the results. Write regular reporting into the contract. Meanwhile, while oversight wobbles, how fast are the tools that take over our work spreading?
CH 02 Agent products line up
On the very day oversight was questioned, tools that act for us only grew in number.Every tool here is pitched as running without a human at the wheel. It learns, repairs and gathers what it needs by itself. The more convenient it gets, the fewer chances people have to step in. So the oversight problem will grow, not shrink. Checking is not keeping pace with delegation.Big tech leaders expect the doorway to information itself to become a delegated assistant. In buying and daily tasks, the screens people operate may shrink. Choices then happen inside the machine, and the reasons become hard to see.Such a system picks its own path. So even a right answer needs a separate check on the route. In pharma, the trail matters as much as the conclusion.Soon these tools will draft literature searches and replies to inquiries. Staff will need rules on what gets logged, more than a user manual. Decide who signs off before you start. But rules on which tools we may use are not only internal. Sometimes a government chooses first.
CH 03 Funding right after a usage ban
Sometimes a government decides first which AI may be used. What happened to the developer it shut out?The order covered AI used inside federal agencies. It told them to drop one developer's technology for security reasons. Yet within days, that developer reportedly pulled in a vast sum. In this story, exclusion did not slow it down.One government's order binds only what its own offices use. It cannot reach where a developer raises money, or who uses it worldwide.The sum is said to rank among the largest in its home country. To me, faith in its momentum outweighed the order. Investors seem to treat the ban as a local hurdle.Drug makers use AI across borders too. A tool allowed at one site may be barred at another. More and more, local rules will decide which AI you may use, before performance does. Map which orders could touch your tools before rollout. If money flows even to a developer that was shut out, investment in AI remains strong. Next, the money closest to pharma.
CH 04 Drug discovery AI: what the sums mean
Some large sums near pharma were reported too. Whether to take them at face value is another matter.At a glance, these sums look alike. They are not. Money already paid and a ceiling paid only if conditions are met mean very different things. A public grant is money that will arrive. Rank deals by size alone, and you misjudge their weight.Drug discovery partnerships often pay in steps, each time development reaches a preset point. If development stops, the rest never comes. So the big total here is best read as a ceiling, not cash in hand.Today's news was only that the first step was passed. Passing it guarantees nothing about the next. We need the habit of separating reported sums from money that moved.Pharma staff often explain partnership news inside their company. Quote only the headline sum, and expectations swell. Add one line on what kind of money it is. That small habit keeps colleagues from misjudging. So what is moving around money and AI right next to the daily work of pharma teams?
CH 05 On the pharma front line
Drug makers are reportedly teaming up with firms that own vast computing power, rather than building AI in house.The partners' work reaches beyond the lab, into running human trials and helping patients. AI is entering the base that runs development itself. The more of that base you hand over, the more the partner's safety and oversight become your own quality. That is why you check their oversight first.The same day brought a bold claim that AI outdiagnoses physicians. A developer was sued over training material used without consent. And written AI rules may help when policies come up for review. Different stories, one question: can you explain yourself?
Pharma has always earned trust through records and explanation. If you cannot show what an AI learned from and how it is managed, the user shares the blame. Bold claims need their evidence checked. Which patients, compared with whom?
It is the same question we ask of a drug. Yet all of this assumes that the electricity and cash behind the computing keep flowing.
CH 06 Power and money as limits
Finally, let us look at the ground under AI's endless growth.One state is defending a halt to new computing sites, citing fears of running short of electricity. Others are building their own supply. Electricity is becoming one of the first things AI growth runs out of. Demand is strong, yet the cash behind it looks thin. A company that cut big investments saw its stock drop.These new compute renters spend heavily just to buy machines. If cash gets harder to raise, their capacity grows slowly. For a renter, today's terms may not hold tomorrow. Plan for price rises, or a provider pulling out.If drug makers move research onto AI, estimate how long that base will last. A partner's electricity and cash troubles become weak spots in your own plans. I would prepare an exit route from day one. Keep data in a format another provider can run. Oversight, exclusion, cash and electricity. Every story today led back to one point: what to check before you hand over the work.
Wrap-up
The work we hand to AI will keep growing. Do not rely only on the maker's word. Ask for outside review and for records. That protects the user. The real question is less what an AI can do, and more who watches it, and how. Next, I will watch how the criticized side responds. That response should show whether it deserves our trust.
- CH 01Reuters「AI safety group METR finds flaws in OpenAI’s internal monitoring system」 reuters.com
- CH 01Mashable「Sam Altman steps down as head of OpenAI's safety group」 mashable.com
- CH 01StateScoop「OpenAI urges Newsom to strengthen California’s AI safety law with additional safeguards」 statescoop.com
- CH 02Google Cloud「Knowledge Catalog: self-learning context enrichment for AI agents」 cloud.google.com
- CH 02Google for Developers Blog「Architecting Governed, Self-Evolving AI Agents on Gemini Agent Platform」 developers.googleblog.com
- CH 02The New York Times「Anthropic Joins the A.I. Agent Race With Claude Duet」 nytimes.com
- CH 02Search Engine Journal「Google CEO Sundar Pichai Views Search Absorbed Into Agentic Layer」 searchenginejournal.com
- CH 03TechCrunch「DeepSeek reportedly raises $15B days after US restrictions」 techcrunch.com
- CH 04BioPharma APAC「Insilico Medicine Hits First AI-Driven Oncology Milestone in $888 Million Servier Collaboration」 biopharmaapac.com
- CH 04Pulse 2.0「Sinopia Biosciences Secures $2.5 Million NIH Grant For AI Metabolomics Drug Discovery Model」 pulse2.com
- CH 04Fortune「We just raised another $550 million to bring AI to the legal industry. I still believe there will be more work than lawyers」 fortune.com
- CH 04Yahoo Finance「How Meta Is Raising $145B for AI Without Giving Up a Single Vote of Mark Zuckerberg’s Control」 finance.yahoo.com
- CH 05The Verge「Microsoft says its AI tool is better than doctors at diagnosing diseases」 theverge.com
- CH 05Bloomberg Law News「Anthropic Sued Over Training Claude AI Model on Protected Music」 news.bloomberglaw.com
- CH 05Law.com「AI Governance Takes Center Stage in Insurance Policy Renewals」 law.com
- CH 06Crusoe「Crusoe and NBSF Announce $700M AI Factory Power Facility」 crusoe.ai
- CH 06FOX 5 New York「Hochul defends New York pause on AI data centers over power concerns」 fox5ny.com
- CH 06CNBC「Nvidia CEO Jensen Huang says neocloud demand is 'off the charts' but capital is scarce」 cnbc.com
- CH 06Reuters「Tesla shares slip after capex cut raises doubts over Musk's AI and robotics push」 reuters.com
Articles used
- AI Daily News 2026-10-23
- AI Daily News, Pharmacy Edition 2026-10-23 Morning
- AI and the Pharmaceutical Industry 2026-10-23
Today's related reports
- AI Daily News October 23, 2026
- AI & Economy News October 23, 2026
- AI & Finance News October 23, 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.