The original is on the right. If Pretti had been Black, would this image "enhancement" have changed his phone into a gun?
Such errors create false narratives *and* manufacture permission to reject photo and video evidence of reality. Both serve the interests of tyranny.
Some aren't even intentional. Tony Thomas is a retired 4-star general and this image was not intentionally manipulated, just "enhanced" from a somewhat blurry frame of a video. It elides the kneeling officer's head, but nobody noticed.
University administrators take note.
Round white badge with slightly left-tilted black sans serif text reading "NOT AI".
Profile badge for @ghostdoc2026.bsky.social
Ghost in the Machine flyer. FREE FILM SCREENING Tuesday, April 14 6-8pm in DUAN G130 RSVP: https://forms.gle/Qb6FsGFWwJEQ2Ygb7 After the screening: stay for an option Q&A about CU's contract with OpenAI, shared governance, and local legislative and labor context.
Boulder folks: come to a free public screening of Ghost in the Machine @ghostdoc2026.bsky.social Tuesday, Apr 14 at 6pm.
This film gives essential context for our profession and invokes research by 🤩 colleagues.
Thanks to CEAS Inclusive Culture Council, @ucwcolorado.bsky.social, and AAUP Boulder.
I've been quietly removing tech from my school library since I took over 6 years ago. I thought people would be annoyed, but it's been great. Adults have barely noticed, and the students have responded well.
I'm one of those trans women. I worked ground ops for Artemis I.
I am torn between celebrating the success of my friends and peers and grieving how my dream of working for NASA was shattered by hateful assholes in Tallahassee
xAI would like courts to hold simultaneously that:
1. illegal outputs (CSAM, unlicensed practice of medicine/law, fraud) from their product are the sole liability of the prompter and
2. outputs are also 1A-protected speech by xAI and cannot be restricted by laws
Our PhD student Kennedy Orwa, who studies applications of AI to health care, was hastily deported today to Kenya along with his 13-year-old son without opportunity to speak to legal counsel.
King 5 reports that he held a valid visa that was rescinded without explanation.
They also decided to conduct a climate survey, but voted to expand the nondiscrimination policy when all they had were vendor quotes for the survey. Meanwhile, the Visiting Scholar in Conservative Thought (which John Eastman was during the insurrection) has only hired conservatives.
Great topic! Local history that may be relevant to your project: regents added political affiliation and political philosophy to the nondiscrimination clause in 2013 after Mark Bauerlein grieved that he didn't feel invited to extend his one-year visiting appointment. www.colorado.edu/today/2013/0...
This applies also domestically: universities, law firms, hospital systems, cities, and states. Courage is contagious.
We're on strike! Don't visit propublica.org on April 8
We’re on strike today! Support our fight for a fair contract by NOT visiting the @propublica.org website or engaging with ProPublica stories today.
Tell ProPublica’s management you won’t cross the picket line: actionnetwork.org/petitions/te...
I had the chance to go on the Boulder Valley Frequency podcast to express my...disgust, I think is the right word, with the contract that CU recently signed with Open AI and with the prospect of integrating ChatGPT throughout higher education #ai www.boulderfrequency.com/2600535/epis...
WE DID IT!
Our illegal union @uam-umd.bsky.social is a lot less illegal today. 12,000 contingent faculty in MD just won the right to collectively bargain. Life-changing win for teachers and students. Only happened bc thousands of people worked together to beat our bosses in the statehouse.
Very interesting paper. Note that any positive effect with assistance is weak/negligible for the reading comprehension task.
Consider the widespread (and likely present here) lack of construct validity in the AI-assisted condition: you get "right" answers for wrong reasons during the assisted phase
“Health economists warned that this AI coding arms race — fueled by AI scribes and autonomous coding tools maximizing codes on one side, and by insurer algorithms trying to minimize payments on the other — is a zero-sum game that could really hurt some of the most vulnerable providers.”
We also don't talk enough about the increased verbosity and reduction in quality of clinical notes. Providers that used to write concise, accurate notes now generate verbose slop that takes more time to pre-chart and has less salience. Same pathologies and self-assessment biases as vibe-coding.
Dai was more worried about the downstream effects of the coding war. What if, as in the recent study from Blue Cross Blue Shield, insurers demand to see treatment to affirm that a diagnosis affixed to a bill is legitimate, and that prompts clinicians to provide inappropriate care to justify the higher bill? What if a primary care physician who might not otherwise have referred a person to a specialist refers them, triggering more unnecessary testing and procedures and preventing people who need that specialist from being able to book an appointment?
The article doesn't touch the legal aspect, but these products also generate errors that go undetected. It could be a clinically-irrelevant comment that generates a new billing code. The provider did not apply the standard of care for that diagnosis, but only the patient can know that.
Figure 1: Extraction of Harry Potter and the Sorcerer’s Stone for a single run. We quantify the proportion of the ground-truth book that appears in a production LLM’s generated text using a block-based, greedy approximation of longest common substring (nv-recall, Equation 7). This metric only counts sufficiently long, contiguous spans of near-verbatim text, for which we can conservatively claim extraction of training data (Section 3.3). We extract nearly all of Harry Potter and the Sorcerer’s Stone from jailbroken Claude 3.7 Sonnet (BoN N = 258, nv-recall = 95.8%). GPT-4.1 requires more jailbreaking attempts (N = 5179) and refuses to continue after reaching the end of the first chapter; the generated text has nv-recall = 4.0% with the full book. We extract substantial proportions of the book from Gemini 2.5 Pro and Grok 3 (76.8% and 70.3%, respectively), and notably do not need to jailbreak them to do so (N = 0). Note: We do not claim we maximized possible extraction for each LLM. Different runs use different underlying generation configurations per LLM.
I agree with unknown revenue stream when the project was started in 2004, but are we to believe that Google does not use Google Books to access the books that its LLM substantially reproduces?
arxiv.org/abs/2601.02671
Day of Action Event Roundup /🧵
Join us on April 16 and 17 as we fight back against Trump's attacks on higher education!
We have a packed slate of organizing workshops and conversations. Bring a friend, bring a colleague.
Find all the details and registration links below ⬇️
How can a teacher assess their own expertise and evaluation? Like nicotine, amphetamines, and gambling, who assesses what constitutes responsible use?
It is unknown what specific practices are sufficient to prevent Rule 11 violations except to ensure the metacognitive processes are performed.
New error modalities flourish and patterns of addiction emerge. The cognitive biases and maladaptive coping are known in theory, yet everyone believes themselves to be immune, while credulously trusting self-assessment. This is a useful post from a co-founder of an AI company.
This is not meant personally, but a lot of critical observers have watched colleagues rave about how the products merely enhance their work, how they're being responsible, still doing all the thinking and carefully checking, etc. And yet quality is slipping relative to their prior work.
AP embraces epistemicide, refusing to bargain with the union over "AI". An organization that refuses to invest in people cannot be a truth-seeking organization.
I’m watching the new essential film, Ghosts in the Machine, featuring many esteemed colleagues of mine, and noting how many of the great men scientists making claims about the ability for machines to mimic humanity also denied the humanity of women in myriad ways.
Thanks, I think we agree on a lot.
I question the premise of having a "leg up". This depends greatly on pedagogy and whether we're referring to ease with which one produces artifacts/grades or to any holistic sense of learning. In general, circumventing friction is circumventing learning.
Categories of deception used in research • Outright deception: Subjects know they are participating in research but are not informed of use of deception. • Covert deception: Subjects are unaware they are in a study, but their environment is manipulated such that the subject has a false understanding of what is occurring. • Incomplete disclosure: Information such as the real purpose of study or the nature of procedures is withheld. • Authorized deception: Subjects are told that they will be misled regarding some aspect of the study.
Relevant regulations: The Common Rule does not discuss research involving deception, but the Belmont Report notes: “In all cases of research involving incomplete disclosure, such research is justified only if it is clear that 1. incomplete disclosure is truly necessary to accomplish the goals of the research, 2. there are no undisclosed risks to subjects that are more than minimal, and 3. there is an adequate plan for debriefing subjects, when appropriate . . .”
Yikes, The Big Hoax is covertly-deceptive human subjects research without an IRB. Participants don't know they're being studied, there is no debriefing, no domain-specificity to the deception, etc.
irbo.nih.gov/documents/14...
See also doi.org/10.1007/s116...
13.2. By Customer Solely to the extent permitted by applicable laws, Customer agrees to indemnify, defend, and hold OpenAI and its affiliates and licensors harmless against any liabilities, damages, and costs (including reasonable attorneys’ fees) payable to a third party arising out of a Claim related to: (a) use of the Services in violation of this Agreement; (b) Customer Applications, if any; or (c) Customer Content. University shall be responsible for the negligent acts and omissions of its officers, agents, employees and representatives with respect to its obligations under this Agreement.
Education (research) *about* smoking does not involve students (or researchers) smoking, even if some do smoke on their own time.
We certainly don't need to buy subscriptions for students when we teach critical literacy, and we don't need schools indemnifying "AI" companies over student use.
What is "critical LLM engagement"?
We need to distinguish between these products as objects of study (what they are/how they work/their externalities and socioeconomic context) versus as methods of inquiry/production (which lack construct validity for most tasks non-experts believe they can do).