October 31, 2026Explainer·10:07·Synthetic narration

The day a proposal refused to bail out AI companies── Who bears losses and liability, filtering before training, broken public models

SPEED

Length 10:07At 1.25× 8:06Chapters 8Size 13 MB

Intro0:37

When an AI investment or an AI tool goes wrong, where do the loss and the blame finally land?Set today's events side by side, and each one points to that single question for me. Lost money, hurt people and poorly made machines do not vanish. Someone has to take them on. For anyone using AI in drug work, that someone may be your own company. Big sums and famous names make people feel safe. But did that comfort come from checking what is inside?

We will move from the large picture down to your own files, looking for what you can verify yourself.

Contents

0:0010:07

CH 011:02

01Who pays if it collapses

Source　Law Commentary ／ Business Insider ／ Moneycontrol.com ／ TradingView

Growing investment, refused bailouts

Those talking investment

Who: AI giants, Foxconn's chairman, Dan Ives
Figure: AI giants' capex this year projected at $828 billion
Statements: Foxconn dismisses bubble fears; Ives sees 2027 as a turning point

Those asking who pays

Who: US House Democrats, including Reps. Casar and Waters
Proposal: No taxpayer bailouts for AI firms or related financial institutions
Position: Failing firms enter bankruptcy; lawbreakers face legal liability

Capex totals Amazon, Alphabet, Meta, Microsoft and SpaceX. Reported by Business Insider

Compare the two sides on screen, and they are not debating the same thing. One speaks about whether growth is coming. The other speaks about what happens if it does not. As I see it, the core issue is who carries the cost of failure. However bright the forecast, it does not answer that.

This is still only a proposal, and it may not pass. Even so, I think it matters that politicians have raised the question openly. Companies building AI into their work must now consider a vendor's decline. If a supplier fails, what happens to the contracted service and the data you entrusted to it? You can ask that at the contract stage, before any trouble starts.

Lost money can be expressed in figures. But when a life or the fruit of someone's work is damaged, where does responsibility go?

CH 021:16

02Lawsuits against AI developers

Source　Courthouse News ／ Netflix Junkie ／ Engadget ／ Just Security

Table 1　Litigation around OpenAI

Plaintiff or commenter	Content
Widow of a man who died	Sued, blaming ChatGPT for her husband's death
Actors', writers' and musicians' unions	Asked the court to reject the fair use claim
Just Security (commentary)	Three changes to make AI firms liable; says current law falls short
Foundation that runs Wikipedia	Weighed in on NYT case: extreme rulings could threaten the open web

The death suit is known only from a Courthouse News headline. Copyright disputes include the NYT's case against OpenAI

These voices come from very different places, yet their question has the same form. How far must a developer answer for what its product does, and for what it learned from? One caution. The claim about a life is known here only from a headline, and no report says a court has ruled.

In disputes over creative work, this idea decides how far use without consent is allowed. If one side wins too much, some warn, people who rely on public knowledge could suffer too. A total victory for either side would not settle it.

The closest case for pharma professionals is feeding papers and internal documents to an AI. If permission is unclear, the developer's dispute could become your company's problem. You need not wait for a ruling. You can write down today what may be entered.

And once a case reaches court, developers are not the only ones questioned. What stays in the user's hands can also become material for assigning blame.

CH 031:07

03AI prompts become evidence

Source　The National Law Review ／ JD Supra ／ FedScoop

Now we turn to daily work. Text typed into an AI does not vanish when you finish. It can remain as a record of your work. That changes what routine use means.

Figure 1　AI records and litigation

Prompts, outputs and conversation histories may have to be produced in discovery. Once litigation is reasonably anticipated, relevant records may need to be preserved.

What concerns me most here is the uneventful period in the middle. Normally, logs just accumulate. Only when a dispute appears does anyone ask what may be deleted and what may not. Deciding at that moment is too late.

Once you know the other side may read it, you write differently. A line typed casually by one employee may be read as a company document.

Table 2　Moves questioning AI records

Forum	Records in question
Corporate litigation	AI prompts, outputs, conversation histories
Corporate discovery	Ransomware logs and Copilot memory
US House committees	AI use in military operations; AI document retention policies

Ranking minority members of two US House committees requested a private briefing

Scrutiny now reportedly reaches beyond human conversation, to what an AI remembers by itself and the traces it leaves as it acts. Governments face the same question as companies.

People who use AI for safety consultations or document drafts should think hard about this. Unpublished data and doubts about a decision go straight into text. Procedures differ by country, so check with your legal team. Then you can decide in advance what to keep, and for how long.

Records get questioned after a problem occurs. So can the builders prevent trouble before it starts?

CH 041:08

04Filtering before training

Source　Tech Times ／ The Cryptonomist

Assigning blame afterward does not prevent harm. New research tries to remove causes earlier. Rather than testing the finished product, it casts doubt on the material at the start.

Figure 2　Filtering before fine-tuning

Reported by researchers at Meta AI, Carnegie Mellon University and the University of Washington. Before fine-tuning, they found seemingly harmless training examples that erode safety, and filtered the data.

The surprise is that innocent-looking content can be the cause. If a document that reads well can still weaken protective behavior, reading it by eye is not enough.

If you adjust a model with your own company's data, this becomes your responsibility. Even a careful developer may not guarantee protection after you have modified it.

A benefit was reported, but the attacks have not vanished. Reducing a risk and removing it are different things. The report does not tell us the test conditions.

Table 3　Another detection report

Item	WM-Basin
Target	Detecting AI models made by distillation
Method	Black-box analysis
True positive rate	79%
False positive rate	0.1%

Reported by The Cryptonomist

The other report tries to tell from the outside where an AI came from. Working from its rates, it misses about one case in five, but false suspicion is very rare. I read it as a useful clue, not as proof.

When choosing a product, ask the same thing. If a vendor claims protective measures, ask what was reduced, by how much, and what remains.

An effect figure means little without its conditions. The same thing is happening in the money stories closest to medicine.

CH 051:03

05What AI drug discovery figures mean

Source　Dealroom ／ Moomoo ／ Tech in Asia

The same habit of checking applies when reading headlines. A big sum makes a company look deeply committed. But that impression has not been verified.

Table 4　Three pharma and AI items

Item	Amount	Nature of the amount
Major pharma and Absci AI antibody discovery deal	Up to $247M	A ceiling, not a confirmed amount
Tempus AI 10% shareholder's common stock sale	US$17.59 Million	Shareholder's sale value; not a company investment
XtalPi's expanded partnership in Singapore	No amount reported	Partner not named in the headline

Actual payments may depend on milestone achievement and other factors

Read down that column, and unlike things sit side by side. One is a future, conditional sum. One is what an owner received for selling shares. One story has no sum at all. None reads as money a company is spending now, so adding them up means nothing.

When payment is tied to progress, a stalled project may mean the rest is never paid. A large sum describes what happens if success continues.

So this figure is not what was paid. It is the most that could be paid. This report does not show how much moves at the start.

When you cite such sums in internal papers, one added phrase about their nature prevents misreading. That matters even more when you brief executives on competitors. A sum's size can distort your own company's investment decisions.

The apparent sum and its substance differed. So what about the AI model that actually reaches your hands?

CH 061:15

06Distributed models that do not work

Source　arXiv

Some believe that running AI on your own servers, inside the company, is safe. That makes sense for protecting secrets. Whether what you obtained can do the job is another question.

Table 5　Test of official Ollama tags

Item	Result
Scope	305 official tags for coding
Failures	7 tags; could solve almost no coding tasks
Failure range	Low-bit quantizations of models with up to 4 billion parameters
Rebuild	5 suspect tags matched rebuilds from the original weights
Calibration added	3 builds of one coding model went from 0 solved to more solved tasks

arXiv preprint, not peer reviewed. Abstract only. Only coding tasks were measured

What strikes me is that some were neither damaged nor fake, yet could not do the work. Building them again gave the same thing, so a delivery accident seems unlikely.

A smaller size makes a model easier to run on local machines. But small models can lose their ability. Loading still finishes normally, so you cannot tell by looking.

A small change in the process produced tasks it could answer correctly. Yet the authors also say the build method alone cannot predict failure.

①

Identity check

Compare tags from the official source with the distributor's values. This shows only that you received the same file that was distributed.

②

Function check

Have it solve tasks close to the planned work; score it against a high-precision version of that model. This shows fitness for those tasks.

③

Authors' proposal

Before release, put low-bit tags through a short functional test. Calibrate those that fail, or hold back their release.

You need to verify two different things. One asks whether you got the same thing. The other asks whether it can work. Passing the first does not answer the second.

Other researchers have not yet confirmed this report, and it measured programming work alone. We do not know whether document summaries behave the same way. So try examples close to your own work before adoption.

Checking a name or an origin does not reveal the quality of the work. So how far has this technology reached into medicines work today?

CH 071:11

07In pharma practice

Source　Bioengineer.org ／ BioSpace ／ ADWEEK

The questions so far are not remote from daily practice. AI stories now stretch from the back of the lab to the patient's hand. Where AI enters decides what gets hurt when it fails.

①

Determining structures

A cryo-electron microscopy review detailed deep-learning automation, from particle picking to high-resolution structure determination.

②

Running experiments

Opentrons raised $40M in a Series E round. It says it will standardize AI-driven science on robots.

③

Connecting with patients

Zocdoc, an app linking medical providers and users, will allow AI agents to access its database.

These examples sit at separate points in medicines work. One studies the shape of molecules. One does hands-on testing. One is where patients meet care. The events differ too: a literature summary, fundraising and a policy statement. None reports that results were proven in practice.

If work on this instrument becomes automatic, human judgment moves to the machine. Results may come faster. But someone must still review the decisions made along the way.

The same goes for the lab and the patient contact point. Is there a record of the steps handed to the machine? When an error appears, who can notice it, and when? I think any meeting on a new system should spend as much time on this as on convenience. In regulated work, steps that cannot be explained should be hard to use.

If work moves to machines, what happens to people's jobs? Here too, how you read the numbers matters.

CH 081:05

08Different numbers on job cuts

Source　JournalArta ／ HR Katha ／ BigGo ファイナンス／ 6abc Philadelphia

Talk of vanishing jobs stirs strong emotions. So numbers tend to spread before anyone checks them. Finally, let us look at how solid the numbers themselves are.

Table 6　Employment reports from the same day

Who	What was reported
NatWest Group	Cut about 1,000 full-time staff in the three months to September
ServiceNow	Explains falling headcount as the result of AI productivity gains
Lao Gan Ma	Rumored cut from about 600 to 48 people through AI inspection
Survey of Korean companies	18% of AI-adopting companies reduced staff
Radiologists	Predicted obsolete 10 years ago; demand now reported at a record high

NatWest's cuts came under a plan to simplify operations. The Korean survey, by Remember & Company, covered 4 industries

Set side by side, these stories differ in how reliable they are. One is a company's own account. One is talk without support. One is a cut made inside a wider business review. In a survey of limited sectors, only a minority cut staff. And a job once declared doomed is reportedly in demand.

We cannot yet say in one line whether AI reduces staff or not. When forecasts fail, I think the people working there lose first. Company explanations alone do not show how large that loss is.

If staffing comes up at your company, ask who gave the number and which work it covers. Drug work likely has many steps built on human review, so figures from other industries do not transfer directly. Measure your own work before setting a department plan.

Even the job figures left the question of who was speaking. Now we return to where we began.

Wrap-up0:19

Watch for how Congress responds to the no-bailout proposal and how markets react to AI giants' capex.

Transcript

Open the full transcript

Intro

We will move from the large picture down to your own files, looking for what you can verify yourself.

CH 01　Who pays if it collapses

The story begins with money. When this much of it pours into AI, people start to talk about cleaning up after a failure. The issue is not whether AI will make a profit. The issue is who pays, and who answers for it, if things go badly.Compare the two sides on screen, and they are not debating the same thing. One speaks about whether growth is coming. The other speaks about what happens if it does not. As I see it, the core issue is who carries the cost of failure. However bright the forecast, it does not answer that.This is still only a proposal, and it may not pass. Even so, I think it matters that politicians have raised the question openly. Companies building AI into their work must now consider a vendor's decline. If a supplier fails, what happens to the contracted service and the data you entrusted to it?

You can ask that at the contract stage, before any trouble starts. Lost money can be expressed in figures. But when a life or the fruit of someone's work is damaged, where does responsibility go?

CH 02　Lawsuits against AI developers

After money comes human harm. Court battles can look like a matter for legal specialists alone. They are not. Their outcomes may shape how responsibility is split between those who build AI and those who use it.These voices come from very different places, yet their question has the same form. How far must a developer answer for what its product does, and for what it learned from?

One caution. The claim about a life is known here only from a headline, and no report says a court has ruled.In disputes over creative work, this idea decides how far use without consent is allowed. If one side wins too much, some warn, people who rely on public knowledge could suffer too. A total victory for either side would not settle it.The closest case for pharma professionals is feeding papers and internal documents to an AI. If permission is unclear, the developer's dispute could become your company's problem. You need not wait for a ruling. You can write down today what may be entered. And once a case reaches court, developers are not the only ones questioned. What stays in the user's hands can also become material for assigning blame.

CH 03　AI prompts become evidence

Now we turn to daily work. Text typed into an AI does not vanish when you finish. It can remain as a record of your work. That changes what routine use means.What concerns me most here is the uneventful period in the middle. Normally, logs just accumulate. Only when a dispute appears does anyone ask what may be deleted and what may not. Deciding at that moment is too late.Once you know the other side may read it, you write differently. A line typed casually by one employee may be read as a company document.Scrutiny now reportedly reaches beyond human conversation, to what an AI remembers by itself and the traces it leaves as it acts. Governments face the same question as companies.People who use AI for safety consultations or document drafts should think hard about this. Unpublished data and doubts about a decision go straight into text. Procedures differ by country, so check with your legal team. Then you can decide in advance what to keep, and for how long. Records get questioned after a problem occurs. So can the builders prevent trouble before it starts?

CH 04　Filtering before training

Assigning blame afterward does not prevent harm. New research tries to remove causes earlier. Rather than testing the finished product, it casts doubt on the material at the start.The surprise is that innocent-looking content can be the cause. If a document that reads well can still weaken protective behavior, reading it by eye is not enough.If you adjust a model with your own company's data, this becomes your responsibility. Even a careful developer may not guarantee protection after you have modified it.A benefit was reported, but the attacks have not vanished. Reducing a risk and removing it are different things. The report does not tell us the test conditions.The other report tries to tell from the outside where an AI came from. Working from its rates, it misses about one case in five, but false suspicion is very rare. I read it as a useful clue, not as proof.When choosing a product, ask the same thing. If a vendor claims protective measures, ask what was reduced, by how much, and what remains. An effect figure means little without its conditions. The same thing is happening in the money stories closest to medicine.

CH 05　What AI drug discovery figures mean

The same habit of checking applies when reading headlines. A big sum makes a company look deeply committed. But that impression has not been verified.Read down that column, and unlike things sit side by side. One is a future, conditional sum. One is what an owner received for selling shares. One story has no sum at all. None reads as money a company is spending now, so adding them up means nothing.When payment is tied to progress, a stalled project may mean the rest is never paid. A large sum describes what happens if success continues.So this figure is not what was paid. It is the most that could be paid. This report does not show how much moves at the start.When you cite such sums in internal papers, one added phrase about their nature prevents misreading. That matters even more when you brief executives on competitors. A sum's size can distort your own company's investment decisions. The apparent sum and its substance differed. So what about the AI model that actually reaches your hands?

CH 06　Distributed models that do not work

Some believe that running AI on your own servers, inside the company, is safe. That makes sense for protecting secrets. Whether what you obtained can do the job is another question.What strikes me is that some were neither damaged nor fake, yet could not do the work. Building them again gave the same thing, so a delivery accident seems unlikely.A smaller size makes a model easier to run on local machines. But small models can lose their ability. Loading still finishes normally, so you cannot tell by looking.A small change in the process produced tasks it could answer correctly. Yet the authors also say the build method alone cannot predict failure.You need to verify two different things. One asks whether you got the same thing. The other asks whether it can work. Passing the first does not answer the second.Other researchers have not yet confirmed this report, and it measured programming work alone. We do not know whether document summaries behave the same way. So try examples close to your own work before adoption. Checking a name or an origin does not reveal the quality of the work. So how far has this technology reached into medicines work today?

CH 07　In pharma practice

The questions so far are not remote from daily practice. AI stories now stretch from the back of the lab to the patient's hand. Where AI enters decides what gets hurt when it fails.These examples sit at separate points in medicines work. One studies the shape of molecules. One does hands-on testing. One is where patients meet care. The events differ too: a literature summary, fundraising and a policy statement. None reports that results were proven in practice.If work on this instrument becomes automatic, human judgment moves to the machine. Results may come faster. But someone must still review the decisions made along the way.The same goes for the lab and the patient contact point. Is there a record of the steps handed to the machine?

When an error appears, who can notice it, and when?I think any meeting on a new system should spend as much time on this as on convenience. In regulated work, steps that cannot be explained should be hard to use. If work moves to machines, what happens to people's jobs?

Here too, how you read the numbers matters.

CH 08　Different numbers on job cuts

Talk of vanishing jobs stirs strong emotions. So numbers tend to spread before anyone checks them. Finally, let us look at how solid the numbers themselves are.Set side by side, these stories differ in how reliable they are. One is a company's own account. One is talk without support. One is a cut made inside a wider business review. In a survey of limited sectors, only a minority cut staff. And a job once declared doomed is reportedly in demand.We cannot yet say in one line whether AI reduces staff or not. When forecasts fail, I think the people working there lose first. Company explanations alone do not show how large that loss is.If staffing comes up at your company, ask who gave the number and which work it covers. Drug work likely has many steps built on human review, so figures from other industries do not transfer directly. Measure your own work before setting a department plan. Even the job figures left the question of who was speaking. Now we return to where we began.

Wrap-up

The person who bears the result when things go wrong is closer than you may think. So I do not treat sums, names or origins as grounds for comfort. Keep your records in order, and test on your own work. That is what users can do. Next, we will watch how politics and markets begin to answer this question.

Sources

CH 01Law Commentary「Democrats Push to Prevent Taxpayer Bailouts for AI Companies」 lawcommentary.com
CH 01Business Insider「AI hyperscalers are expected to finish the year spending a whopping $828 billion on capex」 businessinsider.com
CH 01Moneycontrol.com「Foxconn teases new OpenAI deal after Sam Altman meeting, says AI boom is no bubble」 moneycontrol.com
CH 01TradingView「Dan Ives Says AI Disruption Is ‘On Full Display’ In Tech Earnings, Sees 2027 As Inflection Year」 tradingview.com
CH 02Courthouse News「Widow sues OpenAI, blaming ChatGPT for her husband’s death」 courthousenews.com
CH 02Netflix Junkie「“Harms of Unlicensed Training” – SAG-AFTRA, WGA & AFM Urge Judge to Reject OpenAI’s Fair Use Exemption」 netflixjunkie.com
CH 02Engadget「The Wikimedia Foundation Weighs In On The NYT's Case Against OpenAI」 engadget.com
CH 02Just Security「Liability Could Rein In Rogue AI Agents. Current Law Falls Short.」 justsecurity.org
CH 03The National Law Review「Every Prompt Leaves a Record: How Businesses Can Manage Litigation Risk From Everyday AI Use」 natlawreview.com
CH 03JD Supra「Ransomware Logs And Copilot Memory Put AI Agent Records On The eDiscovery Agenda」 jdsupra.com
CH 03FedScoop「Trump’s AI use to inform military operations questioned by House duo」 fedscoop.com
CH 04Tech Times「Meta AI and CMU Researchers Cut AI Jailbreak Rates by 22% With Pre-Deployment Training Data Filter」 techtimes.com
CH 04The Cryptonomist「AI Distilled Model Detection with WM-Basin Technique」 en.cryptonomist.ch
CH 05Dealroom「AstraZeneca signs up to $247M AI antibody discovery deal with Absci」 app.dealroom.co
CH 05Moomoo「Tempus AI(TEM.US) 10% Shareholder Sells US$17.59 Million in Common Stock」 moomoo.com
CH 05Tech in Asia「Chinese AI drug firm XtalPi expands Singapore partnership」 techinasia.com
CH 06arXiv「Broken on Arrival: Non-Functional Quantized LLM Artifacts in Public Model Registries and How to Catch Them」 arxiv.org
CH 07Bioengineer.org「AI Takes Over Cryo-EM: Deep Neural Networks Now Automate Particle Picking,」 bioengineer.org
CH 07BioSpace「Opentrons Announces $40M Series E Round to Standardize AI-Driven Science on Its Robots」 biospace.com
CH 07ADWEEK「Zocdoc Opens Its Platform to AI Agents」 adweek.com
CH 08JournalArta「NatWest Cuts 1,000 Jobs in Push to Become AI-Driven Bank」 journalarta.com
CH 08HR Katha「ServiceNow links AI-driven efficiency to declining workforce; says smaller teams are delivering more」 hrkatha.com
CH 08BigGo ファイナンス「韓国企業、AI相談員導入でも人員削減は18%…「増員・再配置」を選択」 finance.biggo.jp
CH 086abc Philadelphia「AI was supposed to replace radiologists but now they're in more demand than ever | Here's why」 6abc.com

Articles used

Today's related reports

AI Daily News　October 31, 2026
AI & Economy News　October 31, 2026
AI & Finance News　October 31, 2026

About

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.

← All editions