> daily_signal(2026_10_25)
Google edited a privacy promise it kept for years, a trial showed AI lifts junior lawyers then takes the skill back, Pennsylvania moved to unmask deceptive AI ads, and an AI invented a new enzyme.
PickBits Daily Signal · Sunday, October 25, 2026
// tl;dr
- Google added an exception to its promise never to use Gmail, Drive, or Photos content for personalized ads, and the trigger is its personal AI. As of October 16 the policy now says it won't, "unless you choose to make data from these services available to other Google services that support personalized ads." Connecting those apps to the Personal Intelligence feature opens the door; advertisers still never see your raw data, and you can opt out in settings.
- A three-month controlled trial of 133 patent attorneys found Google Labs' InFlow AI lifted junior drafters the most, but when the AI was removed only senior attorneys kept any gain. The juniors who looked most improved retained none of it, suggesting AI can make inexperienced workers look far more capable while skipping the practice that builds durable expertise.
- Pennsylvania passed SB 806, the AI Deceptive Advertising Act, and it now awaits Governor Josh Shapiro. It would require a clear label whenever AI-generated or AI-altered content is used to mislead consumers about a product, with fines up to $3,000 per violation enforced by the AG and county DAs. Plain text, parody, and obvious fiction are exempt.
- Researchers used AI to design de novo enzymes, catalysts that do not exist in nature, for a hard step in making the HIV-prevention drug lenacapavir. Of 67 lab-tested designs, four worked, and the best converted 79% of its target with the required selectivity. It is an early, pre-peer-review proof that AI can invent working manufacturing catalysts.
The one that actually changed something documented this week is the Google edit. We've followed the opt-out-by-default AI land grab all year, from Meta's Instagram-tagging default in July to Google's own search-history-for-training toggle the same month, and this is the quiet next step: a line added to the privacy policy on October 16 that lets your Gmail and Photos feed ad targeting once you switch on its personal AI. Nobody announced it; Android Authority found the diff.
The other three all turn on the same question: who has to tell you what the AI did. A controlled trial of patent attorneys found AI makes juniors look senior and then takes the skill away when it leaves. Pennsylvania passed a law to force deceptive AI-made ads to carry a label and put it on the governor's desk. And a lab used AI to invent an enzyme that does not exist in nature, to make a hard-to-manufacture medicine cheaper. The enzyme one is the outlier in a day mostly about what companies would rather you didn't notice, and it's the one I'd read first.
Google edited a privacy promise it had kept for years the same week Pennsylvania moved to unmask deceptive AI ads and a Manchester lab let an AI invent an enzyme from scratch.
1. Google quietly added an exception to its promise never to use your Gmail or Photos to sell ads.
The switch that opens it is the personal AI the company is urging everyone to turn on.
For years, Google's privacy policy carried a flat promise: it would not use the content of services like Gmail, Drive, and Google Photos to personalize ads. As of October 16, that promise has a clause attached. The policy now reads that it won't, "unless you choose to make data from these services available to other Google services that support personalized ads." The change was spotted by Android Authority and reported by The Verge. The door it opens is a specific one: the Personal Intelligence feature, which lets Search and Gemini read your mail and photos to give you better answers. Connect those apps, and that same content can inform the ads you see.
This is also easy to overstate, so here is the real limit. Advertisers never get your raw emails, your images, or identifying details like your name, and Google says it will not target you on sensitive categories such as race, health, or religion. So this is not Google auctioning your inbox. The trade is quieter than that: the AI gets more helpful, your ads get to see more of you, and the ad half rides along with the feature, with no separate yes, left for you to go find and switch off. The opt-out lives in the Search Services Personalization settings, where you can disconnect individual apps or disable Personalized Recommendations entirely.
Why this matters: Right now, if you have turned on Google's Personal Intelligence, the content of your Gmail and Photos can shape the ads you see, and the policy change that allowed it shipped with no notice. This is the pattern we keep flagging: the change that matters was a line edited into a policy, not a product anyone announced, and it rides on a feature whose fine print now quietly includes ads. The sensitive-category carve-outs are genuine, but they are Google policing itself. The connection you actually control, between your private apps and the ad machine, is one you may have switched on for better answers without realizing it now feeds ads too.
Action this week: Open your Search Services Personalization settings and look at which apps you have connected to Personal Intelligence, then disconnect the ones you don't want priced into ad targeting. I would keep the feature if you actually use it, since it is useful, and just cut the data you don't want feeding ads, because connecting those apps quietly bundled that in. If you administer accounts for a family or a team, check whether the connection is on for people who never chose it.
2. A controlled trial found AI improved junior lawyers' drafting the most while it was switched on.
When the researchers took the AI away, only the seniors kept the gain, and the juniors kept none of it.
A three-month randomized controlled trial, reported through the National Bureau of Economic Research, put Google Labs' InFlow AI assistant in front of 133 patent attorneys across 11 US law firms. While the attorneys had the assistant, the junior lawyers improved the most, their drafting gaining far more than the seniors'. Then the trial removed the AI. The senior attorneys held onto their gains. The juniors held onto none of theirs.
That result inverts the usual pitch for workplace AI, which is that it lifts your weakest people up toward your strongest. This trial says the lift is real but rented: it vanishes when the tool does, and only the people who already carried the expertise could internalize it. The reading that fits the data is uncomfortable, because the junior who looked most improved had, in skill terms, built the least. It lands next to a story we covered two days ago, Anthropic nearly freezing its junior hiring on the logic that AI now does the junior work, and this is the mechanism under that bet, measured rather than assumed.
Why this matters: This is the workplace-AI result I keep coming back to, because it catches the thing every rollout misses. The junior's work got better and the junior did not, and people treat those as the same fact. Your productivity dashboard shows you the first and never the second. So when you hand the tool to the people who most need the reps, the numbers go up while the skill quietly doesn't, and you won't find out until the AI is gone or the senior who was covering for it leaves.
Action this week: Separate two questions in writing before you call an AI rollout a success: is the work better, and is the worker getting better. Those came apart in this trial, and every time I have watched a team measure only output, the bench thinned a year later without anyone deciding it should. Keep the hard reps, the drafting from scratch and the review without the assistant, on the calendar for the people still learning the craft, or you get this quarter's output and a thinner bench a year from now.
PickBits Daily Signal is free, and the best way to support it is to forward it to one person who would read it. Subscribe today!
3. Pennsylvania passed a law to make deceptive AI-made ads carry a visible label.
It is on Governor Shapiro's desk now, with civil fines up to $3,000 a violation.
Pennsylvania's legislature passed SB 806, the AI Deceptive Advertising Act, and it now awaits Governor Josh Shapiro. The House sent it over on October 7, 163 to 40, after the Senate had backed it 48 to 2 in June; the sponsor is state senator Nick Pisciottano. If Shapiro signs, any ad that uses AI-generated or AI-altered content to mislead consumers about a product's nature, origin, or authenticity would have to carry a clear, conspicuous label. Violations draw civil fines of up to $3,000 each, enforced by the attorney general and county district attorneys, and the law would take effect 60 days after it is signed.
The scope is narrower than a blanket "label all AI," and the narrowness is the point. Content that is only text is exempt, and so are obvious parody and fiction; what the bill targets is deception, the ad that uses a synthetic image or clip to make you believe something false about what you are buying. That is a different instinct from the disclosure rules that frame every synthetic pixel as a problem, and it is the version most likely to survive a First Amendment challenge, because it hangs the requirement on consumer harm rather than on the mere use of a tool.
Why this matters: We've watched this disclosure patchwork build state by state, New York requiring AI-ad labels in June and California mandating performer disclosure in September, and Pennsylvania is the entry with the sharpest consumer hook. My own read is that the deception framing is why this one has teeth: the harm it names, paying for something the pixels lied about, is one an ordinary buyer actually feels, which makes it far easier to enforce than a rule that asks whether a shade of sky was retouched. The catch sits in that same word, misleading: proving an ad misled you about authenticity is exactly what a good advertising lawyer gets paid to fight over, so expect the real battles to be about what counts.
Action this week: Call the governor's office if you are in Pennsylvania and say plainly whether you want SB 806 signed, because it is still a live decision rather than settled law. Everyone else, treat an ad that looks too clean to be real as a prompt to check the seller, not the image, and watch your own statehouse, since the state-by-state pattern says a version of this is coming to you next. If you run marketing, assume the strictest of these laws is the one you will have to comply with and label deliberately now.
mediapost.com: Pennsylvania passes bill requiring disclosure of misleading AI ads (October 8, 2026)
ginsburglawgroup.com: Analysis of Pennsylvania's enrolled AI Deceptive Advertising Act, Printer's No. 1808 (October 9, 2026)
pasenate.com: Pisciottano AI transparency bill advances (2026)
4. A lab used an AI to invent an enzyme that doesn't exist in nature, to help make an HIV drug.
Of 67 designs, four worked, early proof AI can build the catalysts hard medicines need.
Researchers at the University of Manchester, working with David Baker's Institute for Protein Design and NVIDIA, used AI protein-design models to create de novo enzymes, catalysts that do not exist in nature, for one of the hardest chemical steps in synthesizing lenacapavir, an HIV-prevention drug. Instead of modifying an existing protein, the models designed brand-new ones aimed at a difficult amide-bond reaction. The team tested 67 candidate designs in the lab; four of them worked, and the best converted 79% of a chiral amine into the target molecule with the selectivity the chemistry required.
This continues a thread we picked up on October 15, when an AI-designed-protein licensing deal between Zealand and Generate showed the field had commercial pull. This is the next rung up and a more concrete one: not a deal on paper, but enzymes that actually ran in a lab and did real pharmaceutical chemistry, with reported design-success rates and speed several times higher than earlier models. The promise, if it holds, is making hard-to-manufacture medicines cheaper, because a bespoke catalyst can replace steps that today demand rare reagents and expensive processing. The honest counterweight is the hit rate: four in 67 is a proof of concept, not a production line, and the work is a preprint that has not yet been peer-reviewed.
Why this matters: David Baker's lab has spent a decade designing proteins that do not occur in nature, and this is one of the clearer cases where one of those designs did load-bearing work in making an actual medicine rather than proving a point in a journal. The claim got more concrete: before, an AI could sketch a plausible molecule; here, a molecule it designed actually ran a reaction chemists find hard. The reason to care is cost, because the expensive medicines are often the ones whose synthesis is brutal, and a custom enzyme that collapses a hard step is the kind of thing that can move a price.
Action this week: Watch for the peer review and for an independent lab reproducing the result before you file this as a manufacturing breakthrough rather than a strong signal. Four working designs out of 67 is a proof of concept, and my read is that it is still a real rung up from the licensing-deal stage we covered on the 15th, because this time the molecules did the job in a flask. If you follow drug pricing, the question to hold is whether a win on synthesis cost ever actually reaches the shelf price, which is a different fight entirely.
singularityhub.com: A new AI model designs enzymes with atomic precision (October 22, 2026)
» What to watch this week
- Whether Governor Shapiro signs SB 806, and whether other states copy its "deception, not disclosure-of-everything" framing. The narrow, harm-based version is the one most likely to survive a court fight, so its wording may become the template others reach for.
- Whether the InFlow trial changes how firms measure AI rollouts, or whether junior hiring keeps freezing. The result argues for tracking skill growth, not just output; the Anthropic hiring move suggests the market may read it the other way.
- Whether Google's opt-out default draws a regulator the way Meta's Instagram-tagging default did in July. A quiet settings change is exactly the kind of move that invited scrutiny last time, and the pattern is now familiar enough to notice.
- Whether the Manchester enzyme work clears peer review and anyone reproduces the 79% result. A preprint is a claim; a reproduced result on someone else's bench is what turns four working designs into a method.
Tomorrow's signal lands here.