Ads beside AI answers: who checks the evidence?── Makers change how they earn, lawmakers warn, and AI runs at home
Download the video (MP4, 14 MB)
Who does an AI tool serve once it starts showing paid messages next to its answers?Today's news is about how the builders of these tools plan to earn money. That shift reaches into pharma work. As the seller's interests grow stronger, the basis for an answer may become harder to see from outside. In our field, we are the ones who must explain the basis for every claim and every figure. We cannot publish a machine's draft as it stands. So who should check that basis?
The builders, the state, or the users?Let's keep that question in mind to the end.
01OpenAI starts showing ads
Source OpenAI ／ TechCrunch
Our story begins with a builder changing the way it earns. When the source of income changes, so does the question of whom the tool tries to please.
Here is how it works. A user asks for a picture, and a seller's message sits right beside the finished picture. What someone asks for hints at which message might work on them. The builder also says it wants to count how well each message works. So user behavior itself becomes a valuable record for sellers. Users are no longer the only ones paying. And a tool tends to serve whoever pays for it.
Pharma staff ask these tools for help every day. How far could those requests feed someone else's business? Before adoption, I would check the contract terms and how inputs are handled. Staff also need to tell a neutral answer from a paid message on screen. That matters even more for free versions used at work. Write these rules into company policy, and people on the ground won't have to guess.
Earning from paid messages is not a need of this builder alone. In the same week, other companies also made news about money.
02IPO signals and a price gap
Source CNN ／ Fortune ／ tech-insider.org
Building these systems means paying huge computing bills, year after year. Who carries that cost, and how? That is the question here.
| Company / subject | What was reported | Source |
|---|---|---|
| Anthropic | IPO expected despite market uncertainty and AI slowdown talk | CNN |
| OpenAI | Mentioned that its IPO has been delayed | Fortune |
| 3 leading models | 20x price gap: GPT-6 Luna, Sonnet 5.5, Grok 4.7 | tech-insider.org |
One company is expected to sell its stock to the public, while another has pushed that step back. With the outlook unclear, views are split. Once listed, a company is watched from outside through its numbers every quarter. Meanwhile, what users pay also differs widely from one builder to the next, according to one comparison. To me, that spread shows that competition among builders has not yet settled.
So when a drug company chooses a tool, it should look at the vendor's finances, not just performance. A change in strategy can suddenly change prices or what you are allowed to use. Avoid depending on a single vendor, and keep room to switch. That is how you stay ready for the long term. This matters most when you rebuild workflows around these tools. Decide before signing how you would handle a price rise or the end of a service.
The more money talk comes into view, the more the builders' stance on safety is questioned. That question reached lawmakers, who called the builders in.
03The New York hearing
Source CNBC ／ Axios ／ ABC7 Eyewitness News ／ Seeking Alpha
Who decides what safe AI means, and where? There is no settled answer yet. That is why words spoken before lawmakers carry so much weight.
- Role
- Researcher who testified at the hearing
- Statement
- Warned of "racing to build and grow our own adversary"
- Role
- Republican side of Congress
- Statement
- Criticized Anthropic's approach as "alarmist"
OpenAI, Anthropic, Google and Meta attended. Ex-Anthropic researcher Coxon also set to testify alongside the giants
Before lawmakers, one expert warned that the builders are creating a danger with their own hands. On that same day, a company known for caution was accused of spreading panic. The same caution looks too weak to some and too much to others. Until outside rules set a standard, judgments will keep swinging by viewpoint. Political winds could even change how builders behave.
In pharma, national rules and formal review define what a safe medicine looks like. For AI, safety is still close to self-reporting by the builders. So users should not rely on builders' words alone. They need their own standards. Which tasks may use AI, and where must a person check? Put that line in writing inside the company. That document lets you explain your choices. Then, however outside opinion swings, your internal standard can stay put.
While the safety debate goes on, the technical distance between nations is quietly closing as well.
04US–China gap at 3%
Source wccftech.com ／ Startup Fortune ／ Inshorts ／ Startup Fortune
When the difference in performance narrows, the reasons for choosing one tool over another change too.
A large language model reportedly ran fast on equipment you could keep at home, not in a company's big data center. Strong performance is becoming usable on your own machines, without an outside service.
The distance between the two nations is also estimated at its smallest ever. One Chinese developer is said to be the main reason.
| Move | What was reported | Source |
|---|---|---|
| DeepSeek V4.1 Flash | Ranked 6th in September | Inshorts |
| Reflection AI | Preparing a US-made open-weight model | Startup Fortune |
That developer's new version also placed near the top of a monthly ranking. In response, an American company is preparing a rival that publishes its inner workings for anyone to run. I see a shift. We are moving from choosing tools by performance to choosing by where they run and who controls them.
Pharma work involves a lot of data that must never leave the company. Patient information and unpublished trial results are examples. An AI that runs on local machines lets you handle such data without sending it out. But which country built the model still needs careful review against company rules.
Now we move from technology to money closer to pharma. There, what we don't know stood out more than what we do.
05The Tempus AI deal
Source timothysykes.com ／ Emerj Artificial Intelligence Research
Money reportedly changed hands between two AI companies close to drug discovery. But the report goes no further than its headline.
| Item | Details |
|---|---|
| Amount | $42M (secured by the listed company in the headline) |
| Counterparty | Tempus AI |
| Nature of the amount | Reported as a fixed sum, from the words "Locks In" |
| Unknowns | Lump sum or staged, direction of funds, use of funds |
| Same-day item | Commentary on practical AI adoption in pharma (Emerj) |
One item selected that day. Category: enterprise AI platform and governance
The wording suggests the sum is settled, not just an estimate. Yet we don't know who pays whom, or what the money is for. We also don't know whether it arrives at once, or in installments as the work hits each goal. From a headline alone, the weight of this deal can't be judged.
A big number can hide the terms behind it. Paid upfront, the same amount is real value. Paid on milestones, it never comes if the goals are missed. Which one it is changes what this deal means.
Pharma professionals need the same care when reading partnership news. Hold your judgment until the payment terms and the direction of money are clear. In internal reports, separate what is known from what is not. And one headline cannot tell you how money flows across the industry.
Keeping the unknown apart from the known also came up, in another form, in research on judging AI answers.
06Research to remove vacuous credit
Source MetaRubric: Learning to Reward for Rubric-Based Reinforcement Learning ／ metarubric: Code for MetaRubric: Learning to Reward for Rubric-Based Reinforcement Learning
When developers train an AI, they sometimes let another AI judge its answers. If the judge is lenient, the trained model grows lenient too.
This study targets one error. The judge rewards content that isn't actually in the answer. If you tweak the question slightly and the score stays put, the judge probably isn't reading. The team then updates the scoring standard itself as training goes on. The rules aren't fixed. They grow along with the model.
Baseline compared
GRPO training with a fixed grader. Neither the rubric nor the grader changes during training.
Results on medical questions
Across several backbones, accuracy on biomedical questions rose versus a fixed grader, the authors report.
What is not shown
The abstract gives no backbones used, no gain on HealthBench-Hard, and no rate of vacuous credit.
On medical questions, the researchers report more correct answers than with the older approach. But the summary doesn't show how the gains differ by model. The work has not yet been peer reviewed.
Pharma will see more cases of one AI checking another. Picture a reply to a medical inquiry. Does it mention patients who should not take the drug? If the judge gives credit for a warning that isn't there, a risky reply passes. People must still spot-check that the supporting text really exists.
Can we trace the basis for an answer? The same question is now moving through everyday tools and rules.
07In pharma practice
Source OpenAI公式 ／ Windows Central ／ ign.com
Can we show the origin of writing or pictures produced by AI? For anyone handling promotional materials, there is no way around this question.
Sources shown in Word
Microsoft's Copilot in Word now reportedly shows the sources of its answers (Windows Central).
EU provenance rules
OpenAI officially published how it will respond to the EU's text provenance rules.
Watermark in game art
Fans found a Google Gemini image watermark in Fortnite's in-game art, and the art was replaced.
A writing tool now points to what its answers are based on. One builder has said how it will meet a European rule on showing where text comes from. Meanwhile, in a popular game, a mark left in an AI-made picture revealed that AI was used. Tools, rules and accidental discovery are all pushing the same way, toward showing origins.
In pharma materials, you can't publish a claim or a figure without explaining its basis. Reviewers ask for that, one item at a time. If an AI draft comes with its references, checking takes less work. If an AI-made figure carries a hidden mark, its use can be traced later. So I recommend logging which steps used AI. Without that log, you have no way to check when reviewers ask.
Behind these stories about records and origins, a few quieter items sat closer to the daily work of users.
08What is being overlooked
Source reuters.com ／ MIXED Reality News ／ Jackson Walker LLP
Some events that got little coverage still bear directly on the work of people who use AI.
Astra trademark suit
Reuters reported that OpenAI was sued for trademark infringement over its AI model name "Astra".
API keys in plain text
An open-source Ray-Ban Meta glasses app kept 8 API keys in plain text. It was not fixed until October 4.
The Westlaw ruling
The Third Circuit affirmed the Westlaw headnotes decision on AI copyright and fair use.
Someone sued, saying the name of an AI system clashes with a name they already own. A tool released for camera glasses left its secret access codes unprotected. And in a fight over using other people's writing to train AI, a higher court upheld the lower court's view. The headlines were small. But they touch the basics of using AI: names, access codes and training material. A small headline does not mean a small impact. Small stories often become big problems later.
When a drug company brings in AI, it always compares performance and price. But name rights, key storage and the rights to training data often slip off the checklist. If a key leaks, outsiders can use AI under your company's account. Rulings on training data also affect who bears responsibility in the vendor contract. Add these checks to your adoption process.
Key storage and rights checks both show one thing. The job of verifying has already moved to the users.
Watching how the giants and the ex-Anthropic researcher explained self-regulation at the NYC hearing, and whether the timing and terms of Anthropic's IPO are revealed.
Open the full transcript
Intro
Who does an AI tool serve once it starts showing paid messages next to its answers?Today's news is about how the builders of these tools plan to earn money. That shift reaches into pharma work. As the seller's interests grow stronger, the basis for an answer may become harder to see from outside. In our field, we are the ones who must explain the basis for every claim and every figure. We cannot publish a machine's draft as it stands. So who should check that basis?
The builders, the state, or the users?Let's keep that question in mind to the end.
CH 01 OpenAI starts showing ads
Our story begins with a builder changing the way it earns. When the source of income changes, so does the question of whom the tool tries to please.Here is how it works. A user asks for a picture, and a seller's message sits right beside the finished picture. What someone asks for hints at which message might work on them. The builder also says it wants to count how well each message works. So user behavior itself becomes a valuable record for sellers. Users are no longer the only ones paying. And a tool tends to serve whoever pays for it.Pharma staff ask these tools for help every day. How far could those requests feed someone else's business?
Before adoption, I would check the contract terms and how inputs are handled. Staff also need to tell a neutral answer from a paid message on screen. That matters even more for free versions used at work. Write these rules into company policy, and people on the ground won't have to guess. Earning from paid messages is not a need of this builder alone. In the same week, other companies also made news about money.
CH 02 IPO signals and a price gap
Building these systems means paying huge computing bills, year after year. Who carries that cost, and how?That is the question here.One company is expected to sell its stock to the public, while another has pushed that step back. With the outlook unclear, views are split. Once listed, a company is watched from outside through its numbers every quarter. Meanwhile, what users pay also differs widely from one builder to the next, according to one comparison. To me, that spread shows that competition among builders has not yet settled.So when a drug company chooses a tool, it should look at the vendor's finances, not just performance. A change in strategy can suddenly change prices or what you are allowed to use. Avoid depending on a single vendor, and keep room to switch. That is how you stay ready for the long term. This matters most when you rebuild workflows around these tools. Decide before signing how you would handle a price rise or the end of a service. The more money talk comes into view, the more the builders' stance on safety is questioned. That question reached lawmakers, who called the builders in.
CH 03 The New York hearing
Who decides what safe AI means, and where?There is no settled answer yet. That is why words spoken before lawmakers carry so much weight.Before lawmakers, one expert warned that the builders are creating a danger with their own hands. On that same day, a company known for caution was accused of spreading panic. The same caution looks too weak to some and too much to others. Until outside rules set a standard, judgments will keep swinging by viewpoint. Political winds could even change how builders behave.In pharma, national rules and formal review define what a safe medicine looks like. For AI, safety is still close to self-reporting by the builders. So users should not rely on builders' words alone. They need their own standards. Which tasks may use AI, and where must a person check?
Put that line in writing inside the company. That document lets you explain your choices. Then, however outside opinion swings, your internal standard can stay put. While the safety debate goes on, the technical distance between nations is quietly closing as well.
CH 04 US–China gap at 3%
When the difference in performance narrows, the reasons for choosing one tool over another change too.A large language model reportedly ran fast on equipment you could keep at home, not in a company's big data center. Strong performance is becoming usable on your own machines, without an outside service.The distance between the two nations is also estimated at its smallest ever. One Chinese developer is said to be the main reason.That developer's new version also placed near the top of a monthly ranking. In response, an American company is preparing a rival that publishes its inner workings for anyone to run. I see a shift. We are moving from choosing tools by performance to choosing by where they run and who controls them.Pharma work involves a lot of data that must never leave the company. Patient information and unpublished trial results are examples. An AI that runs on local machines lets you handle such data without sending it out. But which country built the model still needs careful review against company rules. Now we move from technology to money closer to pharma. There, what we don't know stood out more than what we do.
CH 05 The Tempus AI deal
Money reportedly changed hands between two AI companies close to drug discovery. But the report goes no further than its headline.The wording suggests the sum is settled, not just an estimate. Yet we don't know who pays whom, or what the money is for. We also don't know whether it arrives at once, or in installments as the work hits each goal. From a headline alone, the weight of this deal can't be judged.A big number can hide the terms behind it. Paid upfront, the same amount is real value. Paid on milestones, it never comes if the goals are missed. Which one it is changes what this deal means.Pharma professionals need the same care when reading partnership news. Hold your judgment until the payment terms and the direction of money are clear. In internal reports, separate what is known from what is not. And one headline cannot tell you how money flows across the industry. Keeping the unknown apart from the known also came up, in another form, in research on judging AI answers.
CH 06 Research to remove vacuous credit
When developers train an AI, they sometimes let another AI judge its answers. If the judge is lenient, the trained model grows lenient too.This study targets one error. The judge rewards content that isn't actually in the answer. If you tweak the question slightly and the score stays put, the judge probably isn't reading. The team then updates the scoring standard itself as training goes on. The rules aren't fixed. They grow along with the model.On medical questions, the researchers report more correct answers than with the older approach. But the summary doesn't show how the gains differ by model. The work has not yet been peer reviewed.Pharma will see more cases of one AI checking another. Picture a reply to a medical inquiry. Does it mention patients who should not take the drug?
If the judge gives credit for a warning that isn't there, a risky reply passes. People must still spot-check that the supporting text really exists. Can we trace the basis for an answer?
The same question is now moving through everyday tools and rules.
CH 07 In pharma practice
Can we show the origin of writing or pictures produced by AI?For anyone handling promotional materials, there is no way around this question.A writing tool now points to what its answers are based on. One builder has said how it will meet a European rule on showing where text comes from. Meanwhile, in a popular game, a mark left in an AI-made picture revealed that AI was used. Tools, rules and accidental discovery are all pushing the same way, toward showing origins.In pharma materials, you can't publish a claim or a figure without explaining its basis. Reviewers ask for that, one item at a time. If an AI draft comes with its references, checking takes less work. If an AI-made figure carries a hidden mark, its use can be traced later. So I recommend logging which steps used AI. Without that log, you have no way to check when reviewers ask. Behind these stories about records and origins, a few quieter items sat closer to the daily work of users.
CH 08 What is being overlooked
Some events that got little coverage still bear directly on the work of people who use AI.Someone sued, saying the name of an AI system clashes with a name they already own. A tool released for camera glasses left its secret access codes unprotected. And in a fight over using other people's writing to train AI, a higher court upheld the lower court's view. The headlines were small. But they touch the basics of using AI: names, access codes and training material. A small headline does not mean a small impact. Small stories often become big problems later.When a drug company brings in AI, it always compares performance and price. But name rights, key storage and the rights to training data often slip off the checklist. If a key leaks, outsiders can use AI under your company's account. Rulings on training data also affect who bears responsibility in the vendor contract. Add these checks to your adoption process. Key storage and rights checks both show one thing. The job of verifying has already moved to the users.
Wrap-up
The more ways builders find to earn, the more the work of checking answers shifts to users. So far, neither governments nor lawmakers have taken on that job. It falls to us. We record origins, have people check the judges, and wait until the terms are clear. Next, we will see how the builders described their own rules to lawmakers. That account will help us set our own standards.
- CH 01OpenAI「Building advertising for the way people use AI」 openai.com
- CH 01TechCrunch「OpenAI launches visual ads that appear alongside image generation results」 techcrunch.com
- CH 02CNN「Anthropic expected to IPO despite market uncertainty, AI slowdown calls」 edition.cnn.com
- CH 02Fortune「Can superintelligence ever be controlled? Sam Altman on AI safety, doomsday scenarios, and OpenAI's delayed IPO」 fortune.com
- CH 02tech-insider.org「GPT-6 Luna vs Sonnet 5.5 vs Grok 4.7: 20x Price Gap [2026]」 tech-insider.org
- CH 03CNBC「AI researcher warns 'we are racing to build and grow our own adversary' in NYC hearing」 cnbc.com
- CH 03Axios「GOP senator warns Anthropic of "alarmist" approach to AI」 axios.com
- CH 03ABC7 Eyewitness News「Google, Meta, OpenAI among firms appearing at NYC AI hearing」 abc7ny.com
- CH 03Seeking Alpha「Ex-Anthropic researcher Coxon to testify alongside AI giants at NYC hearing (ANTHRO:Private)」 seekingalpha.com
- CH 04wccftech.com「The 552B DeepSeek V4.1-Flash Model Offers A Peak Output Of 494 Tokens/Second When Powered By An At-Home Rig Spanning 4x NVIDIA DGX Spark Units」 wccftech.com
- CH 04Startup Fortune「DeepSeek Narrows AI Gap With US to Just 3 Percent, Bloomberg Says」 startupfortune.com
- CH 04Inshorts「US-China AI gap narrows to record-low 3%, led by DeepSeek: Report | DeepSeek’s V4.1 Flash ranked 6th in Sept | Inshorts」 inshorts.com
- CH 04Startup Fortune「Reflection AI readies a US open-weight model to challenge DeepSeek and Qwen」 startupfortune.com
- CH 05timothysykes.com「RXRX Rises As Tempus AI Deal Locks In $42M」 timothysykes.com
- CH 05Emerj Artificial Intelligence Research「Accelerating Evidence to Action in Pharma with Practical AI Adoption」 emerj.com
- CH 06MetaRubric: Learning to Reward for Rubric-Based Reinforcement Learning（「必要な情報を回答から取り除いても加点が残ること」）
- CH 06metarubric: Code for MetaRubric: Learning to Reward for Rubric-Based Reinforcement Learning（「採点基準を固定せず、学習の進み具合に合わせて育てるという設計である。」）
- CH 07OpenAI公式「Our approach to EU text provenance rules」 openai.com
- CH 07Windows Central「Copilot in Word finally shows where its answers come from」 windowscentral.com
- CH 07ign.com「Fortnite Has Updated In-Game Art After Fans Discovered a Google Gemini AI Image Watermark」 ign.com
- CH 08reuters.com「OpenAI sued for trademark infringement over 'Astra' AI model」 reuters.com
- CH 08MIXED Reality News「An open-source Ray-Ban Meta glasses app kept eight API keys in plain text until October 4」 mixed-news.com
- CH 08Jackson Walker LLP「Training Day: Third Circuit Affirms Westlaw Headnotes Decision in AI Copyright Fair Use Case」 jw.com
Articles used
- AI Daily News 2026-10-06
- AI and the Pharmaceutical Industry — 2026-10-06
- AI Daily News 2026-10-05
- AI and the Latest Technology: Removing graders that score what an answer never says from reinforcement learning
Today's related reports
- AI Daily News October 6, 2026
- AI & Economy News October 6, 2026
- AI & Finance News October 6, 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.