> daily_signal(2026_08_25)
Who has to answer for AI showed up in four rooms this week: an Arkansas courtroom, New Mexico's Supreme Court, a $40 million AI budget, and an FDA clearance.
PickBits Daily Signal · Tuesday, August 25, 2026
// tl;dr
- xAI is suing the Arkansas photographer arrested after police found roughly 1,700 sexual images and videos of minors, much of it generated with Grok, and it wants him to cover xAI's legal costs. The complaint calls Grok "a powerful, neutral" tool where "every generation is the result of the user's prompts," invoking a terms-of-service indemnification clause, even as it concedes Grok refused the prompts until he jailbroke its guardrails.
- New Mexico's Supreme Court unanimously froze both the air-permit hearing and a new-well water authorization for Oracle's hyperscale Project Jupiter data center near Santa Teresa. The stays came on petitions from the State Engineer, the Center for Biological Diversity, and New Energy Economy. Responses are due September 2, and the DOJ is investigating allegedly fraudulent public comments in the air-permit review.
- Thomson Reuters spent about $40 million over two-plus years building its own AI model, "Thomson," on Alibaba's open-weight Qwen, to stop renting frontier models from OpenAI and Anthropic. Only about $450,000 of that was the final training run; the rest was staff, compute, licensed content, and expert labor. The model now powers citation-checking inside CoCounsel Legal.
- The FDA granted 510(k) clearance to Tempus ECG-PH, which flags signs of pulmonary hypertension from the standard 12-lead ECG most heart patients already get. The condition is badly underdiagnosed because gold-standard testing is invasive, and it affects up to 10% of adults over 65. It is an early-warning aid, not a diagnosis, and shipped with no published accuracy figures.
Four rooms this week, and in each one someone tried to settle who answers for what AI does. xAI, whose "Grok is a neutral tool, the user is liable" posture we watched the Justice Department endorse against Minnesota only last week, is now running that same argument against a photographer who used Grok to make child-abuse images, and asking him to pay xAI's legal bill. A New Mexico court did the reverse, freezing a data center until the public process it tried to skip could catch up. Between the two sits a quieter kind of accountability: a $40 million decision to own an AI rather than rent one, and an FDA clearance that hands clinicians a new flag but none of the data to trust it yet.
None of these is clean. xAI's own filing admits Grok refused before it was jailbroken, which is the fact that turns "neutral tool" from a defense into a question. The data-center freeze is a real brake on a buildout that usually rolls straight over the locals, but it rests partly on a fraud probe that has not concluded. Thomson Reuters' bet on Alibaba's Qwen looks shrewd until you remember Alibaba has already added commercial strings to its "open" weights. And a cleared medical AI is only as trustworthy as the validation nobody has published. The through-line isn't who's right. It's that "who is responsible for this AI" stopped being a rhetorical question this week.
A courtroom, a state Supreme Court, a $40 million budget line, and an FDA review all turned on the same question: who answers for the AI.
1. xAI is suing the man arrested for using Grok to make child-abuse images.
Its filing calls Grok "a powerful, neutral" tool and invokes a terms-of-service clause to make him pay the company's legal costs, even while conceding Grok refused the prompts until he jailbroke it.
There is no settled public rule for who answers when an AI tool produces something illegal, so xAI is trying to write one in a courtroom. After police in Bentonville, Arkansas arrested photographer Russell Bloodworth over roughly 1,700 sexual images and videos of minors, much of it generated with Grok from photos he had taken of juvenile clients, xAI did not defend its own safeguards. It sued him. The complaint calls Grok "a powerful, neutral generative artificial intelligence tool" and argues "every response, every image, every generation is the result of the user's prompts and directions," then invokes an indemnification clause in its terms of service to make Bloodworth pay xAI's legal costs, seeking a declaration that he breached the acceptable-use policy plus an injunction barring new accounts.
The record cuts against the framing. xAI's own filing concedes that Grok initially refused the prompts on content-moderation grounds and that Bloodworth used repeated adversarial prompts to get around the guardrails, which is the detail the Arkansas families already suing xAI keep pointing at. Their attorney, Derek Potts, argues Grok shipped with weaker safeguards than rival platforms. "The user did it" is a cleaner story when your own complaint does not say the guardrails worked until someone broke them. We have tracked this posture since xAI sued Minnesota over its nudification-app ban, and the Justice Department sided with the company; this is the same argument on a far darker set of facts.
Why this matters: When an AI tool produces something illegal, the fine print may already say the account holder, not the company that built the tool, owes the legal bill. That is a contract most people click past, and xAI just showed it will try to enforce it in court. At work, the account holder is your company, which turns a click-through you never read into a liability you now own.
Action this week: Pull up the terms of service for any third-party AI your staff touches and read the indemnification and acceptable-use clauses before an incident, not after. Whenever I have asked a vendor "who is liable if your model outputs something illegal," the answer lives in a clause nobody read, and this suit is what enforcing it looks like. For trust-and-safety and legal teams, log the refuse-then-jailbreak sequence, because that exact record is what both sides are now fighting over.
petapixel.com: xAI sues photographer, blaming him for sexual images created with Grok (August 24, 2026)
katv.com: Elon Musk's xAI suing Bentonville photographer accused of using Grok to generate CSAM (August 2026)
arkansasonline.com: Elon Musk's company's lawsuit blames Bentonville man (August 21, 2026)
2. New Mexico's Supreme Court froze Oracle's giant AI data center over its water and its air.
All five justices paused the air-permit hearing and a new-well water authorization; responses are due September 2, and the DOJ is separately probing faked public comments in the review.
The usual pattern is that a data center gets its permits before anyone can argue the process skipped a step. This time a court restored the step. On August 24 the New Mexico Supreme Court unanimously issued two stays against Project Jupiter, the hyperscale AI data center Oracle is building in Santa Teresa with Santa Teresa Capital, LLC. It paused the state's air-quality permit hearing and froze an emergency authorization that let the project draw water from a new well during construction. The petitioners, the State Engineer alongside the Center for Biological Diversity and clean-energy group New Energy Economy, argued the Environment Department cannot hold an air hearing before the project's fuel source is even settled.
That fuel source keeps moving: Oracle recently swapped a gas-turbine-and-diesel design for Bloom Energy fuel cells, promising lower water use and emissions and pledging carbon-free electricity by 2031. Underneath the permitting fight, the Department of Justice is investigating allegations that fraudulent public comments were filed during the air-permit review. This is the data-center water fight we have followed since residents first worried about a facility draining the desert, and it is a rare moment where the process actually slowed one down. Responses are due September 2, and the stays hold until the court rules.
Why this matters: For once the water under a desert town got a hearing before the data center did. My own read is that the ratepayers and the aquifer just won a round they usually lose, and the fraud probe is the more durable half of the story, because a permitting process is only as good as the comments it counts. These sites fight hardest over the two resources you also need, and a court just said the public process comes first.
Action this week: Watch the September 2 filings and the fuel-source question the court flagged, because whether an air hearing can run before the power source is locked is the precedent here. If a data-center permit window opens near you, verify the provenance of the comments in the docket rather than trusting the tally, which is exactly what the DOJ probe is about. And read the petitions from New Energy Economy for the specific water and emissions arguments that moved all five justices.
kvia.com: New Mexico Supreme Court pauses Project Jupiter water well drilling and air permit proceeding (August 24, 2026)
abqjournal.com: NM Supreme Court pauses Project Jupiter air permit (August 2026)
santafenewmexican.com: New Mexico Supreme Court pauses Project Jupiter permitting processes (August 2026)
3. Thomson Reuters spent $40 million building its own AI so it could stop renting from OpenAI and Anthropic.
Only about $450,000 of that was the final training run; the rest was staff, licensed content, and expert labor on top of Alibaba's open-weight Qwen.
This is the build-versus-rent decision landing on a real budget. On August 24 Thomson Reuters detailed "Thomson," a proprietary AI model it built over more than two years for roughly $40 million, of which only about $450,000 was the final training run. The rest was staff, compute, decades of licensed content, and expert labor. It is based on Alibaba's open-weight Qwen3.5-397B, first retrained with Imperial College for safety and neutrality, then trained on Thomson Reuters' own legal, tax, and compliance content, and it now powers the Tabular Analysis feature in CoCounsel Legal, handling work like citation checking. CTO Joel Hron framed the economics as "renting a house versus buying a house," the point being to escape provider lock-in on inference cost and roadmap rather than fine-tune a frontier model from OpenAI or Anthropic.
The interesting part is where the money actually went. The GPUs were the cheap line; the content and the people were the expensive one, which is the opposite of how most "we trained our own model" pitches are sold. There is a catch worth putting in writing, though: Alibaba has already added commercial restrictions to some Qwen releases, so a model you own on someone else's open weights isn't the same as a model you own outright. This is the clearest case yet of a large incumbent deciding a domain-tuned model on open weights can beat renting a general-purpose frontier one.
Why this matters: We have watched open-weight models chip at the frontier labs' pricing power all year, and Thomson Reuters just put $40 million behind the trade. The number that should reframe your own math is that the training run was only about $450,000 of it, so if you're pricing build versus rent, the GPUs aren't where the cost lives. The payoff they are buying is escaping someone else's inference pricing and roadmap.
Action this week: Price the build-versus-rent question the way they did, and budget for content and expert post-training rather than the training run, because that is where the $40 million actually went. I would test an open-weight base tuned on your own data before fine-tuning a frontier model you cannot control, but read the Qwen license first, since the commercial terms on "open" weights have already tightened. In a regulated domain, copy their sequence and bake in the safety and neutrality training before you pour proprietary content in.
the-decoder.com: Thomson Reuters bets $40M on owning its AI instead of renting from OpenAI or Anthropic (August 24, 2026)
thenewstack.io: Thomson Reuters built its own AI model (August 2026)
PickBits Daily Signal is free, and it is the fastest daily read on where AI power is actually being decided. If it is worth something to you, the best support is to forward it to one person who would read it. Subscribe today!
4. The FDA cleared an AI that flags a deadly, often-missed lung condition from a standard ECG.
Tempus ECG-PH reads the 12-lead ECG most heart patients already get; it is an early-warning aid, not a diagnosis, and it shipped with no published accuracy data.
Start with the person this is for: a patient over 40 with unexplained breathlessness whose pulmonary hypertension goes unnamed for years because the early symptoms are vague and the gold-standard workup, a right-heart catheterization, is invasive. On August 24 Tempus AI received FDA 510(k) clearance for Tempus ECG-PH, which analyzes a standard 12-lead ECG, the cheap and ubiquitous test most cardiac patients already get, for signs of elevated mean pulmonary artery pressure (above 20 mmHg) in patients 40 and older who have cardiac symptoms but no known PH. PH affects roughly 1% of people and up to 10% of adults over 65, so the diagnostic gap is large. The constructive point is that this aims AI at that gap instead of adding a new billable test, and it is Tempus's third cleared cardiovascular AI, after atrial-fibrillation and low-ejection-fraction detection.
The caveat is the same one we raise every time one of these clears. Tempus ECG-PH is explicitly "not intended to be a stand-alone diagnostic tool," and no sensitivity or specificity figures accompanied the clearance. That matters because researchers have flagged for a while that plenty of FDA-cleared AI devices ship without real-world validation, and an early-warning flag that mostly generates false positives makes a clinic's day worse, not better. Cleared isn't the same as validated.
Why this matters: A killer that usually hides behind vague symptoms could get caught earlier, from a test hospitals already run, for the millions of older adults who never get worked up for it. That is a genuinely good use of AI, pointed at a diagnostic gap rather than a new revenue line. The honest limit is that "cleared" tells you the FDA let it on the market, not that it works well in your clinic.
Action this week: Treat this as an early-warning overlay on the ECGs you already run, but demand the sensitivity and specificity data before it touches triage, because none was published with the clearance. Define the downstream pathway now, so a positive flag routes to a real confirmatory workup (echocardiography, right-heart catheterization) and leads to a diagnosis rather than anxiety. And if you're over 40 with breathlessness nobody has explained, it's a fair thing to raise with a clinician.
» What to watch this week
- Whether a court lets xAI's "neutral tool" indemnification argument stand, and how the parallel Arkansas family suits land. The countersuit sidesteps whether Grok was safe enough in the first place; the families' cases put that question directly, and together they will start to define who eats the liability when an AI's guardrails are provably jailbroken.
- The September 2 responses in New Mexico, and whether the DOJ's fake-comments probe produces charges. The stays hold until the court rules, and the deeper precedent is whether an air-quality hearing can proceed before a data center's fuel source is settled, which would give every desert community a new lever.
- Whether other large incumbents follow Thomson Reuters into owning a domain-tuned open-weight model, and how Alibaba's tightening Qwen terms affect them. The build-versus-rent math just got a public data point; the open question is whether "open" weights stay cheap enough to build on once their owners start adding commercial strings.
- Whether Tempus publishes real-world sensitivity and specificity for ECG-PH, and whether clinicians adopt it before those numbers exist. This is the recurring test for FDA-cleared AI: a clearance is a market entry, not a validation, and the useful signal is whether the company backs the flag with the data a clinic needs to trust it.
Tomorrow's signal lands here.