Private beta

The clearest evidence so far comes from medicine. After AI assistance became routine at four endoscopy centers, the doctors' detection rate on the colonoscopies they did without AI fell by about six percentage points, and these were experienced clinicians with thousands of procedures behind them (Budzyn et al., 2025). Nobody decided to get worse. The skill moved into the tool, and it was missing on the day the tool was not there.

In a survey of 319 knowledge workers, the more people trusted the AI, the less critical thinking they reported doing, and the more they trusted their own judgment, the more they did. The work itself moved from producing an answer to verifying and managing what the machine produced (Lee et al., 2025). The output looks the same, and often better, so the shift is easy to miss.

In a field experiment, students given an unguarded AI tutor did better while they had it and then scored worse on their own than students who never had it, while a version built with guardrails avoided the harm (Bastani et al., 2025). AI does not ruin learning by itself. The tool has to be set up so the person still does the thinking, and that matters most for staff who are new and still building judgment.

When the draft, the summary, and the plan all arrive pre-made, the job turns into approving a machine's output. Much of what makes a caseworker good is the thinking that happens while writing the note and the discomfort that sends you back to the file, and that is the part the tool removes first. Say it out loud in supervision. A worker who feels like a reviewer of someone else's work is describing something real.

Do the hard case yourself first, then ask the tool and compare. Give new staff cases without the tool, on purpose, the way any profession trains. Treat "AI drafts, you decide" as a rule that protects your competence and not only your clients' accuracy.

Everything the model has came from text that already existed in a form a company could collect. Knowledge that lives in speech, in a community, in paper files nobody scanned, or in a language with little written presence online is simply absent, and the model cannot notice its own absence. Who had money, publishing, and infrastructure decided what got digitized, so the shape of a model's knowledge follows the shape of economic and colonial history.

Vietnamese, Tagalog, Amharic, and Haitian Creole are each spoken by tens of millions of people. They are called low-resource for one reason: little of that language sits online as clean digitized text. Researchers have sorted the world's languages into tiers on exactly this basis, and the distance between the top tier and the rest is enormous (Joshi et al., 2020).

Where a language is thin online, the corpora get padded with whatever can be scraped, including religious translations, government notices, and machine-translated pages. An audit of the standard web-crawled multilingual datasets found systematic problems in the lower-resource corpora, with several holding less than half acceptable-quality sentences and dozens carrying the wrong language label altogether (Kreutzer et al., 2022). When a model answers your client in Amharic, that can be the material underneath it.

Because the text does not exist, companies pay to create it, scanning, transcribing, translating, recording, and writing the missing language. Meta released a translation model covering 200 languages (Meta AI, No Language Left Behind), and organizations such as Karya pay rural workers in India to record speech in their own languages, then sell the datasets to large technology companies (Karya). Both things are true at once. A language with almost no digital record can gain one, and the people who supplied it are usually paid contract rates, keep no ownership of the resource, and then watch the capability get sold back to them. Consent, pay, and who ends up owning the corpus decide which of the two it is.

A separate stream of work makes the model safe and agreeable, rating answers, ranking which reply is better, and filtering violent and sexual material out of the training data. Most of it happens in English, and it is outsourced to countries where English is widely spoken and wages are low, a pairing that exists because of colonial education systems. Reporting documented Kenyan workers paid on the order of two dollars an hour to read traumatic material so a chatbot would not repeat it (Perrigo, 2023), and peer-reviewed research on data-work platforms in Latin America found the same structure, in which the people who produce the data hold the least power in the chain (Miceli & Posada, 2022).

The published material was largely unlicensed as well. Authors sued Anthropic over millions of books, including copies pulled from pirate libraries and print books the company bought, sliced apart, and scanned (Washington Post, 2026); that case settled for about $1.5 billion. Newspapers led by the New York Times are still litigating against OpenAI (Reuters, 2026). One federal judge has held that training on lawfully obtained books is fair use and that downloading pirated ones is not. Other cases remain open, and the bill is unsettled.

Google measured its own assistant and reported a median text prompt at about 0.24 watt-hours of electricity and 0.26 millilitres of water, roughly nine seconds of television and five drops (Google, 2025). Other estimates for a chatbot query run closer to 2.9 watt-hours, about ten times a web search (MIT Technology Review, 2025). Estimates differ by an order of magnitude, they usually leave out the training run and the manufacture of the hardware, and the companies that could settle the question mostly do not publish.

US data centers used about 4.4 percent of the country's electricity in 2023, and a Department of Energy laboratory projects between 6.7 and 12 percent by 2028, with AI as the main driver (Lawrence Berkeley National Laboratory, 2024). Your one prompt is a rounding error. The demand curve that many prompts add up to is what gets a power plant built.

A reasoning model that thinks for a minute, a long document pushed through a large model, and an image or a video cost far more than a short question to a light model. Beyond speed and money, this is a third reason to leave the light model as the default and reach for the heavy one deliberately. A small model doing one narrow job is the cheapest arrangement on every axis, including this one.

Data centers go up somewhere, and the somewhere is rarely wealthy. In Memphis, a company powered an AI data center with dozens of gas turbines next to a historically Black neighborhood that already carried a heavy pollution burden, and it drew a Clean Air Act suit and organized community opposition (TIME, 2025). In Santiago, an environmental court sent a data center permit back for revision after residents challenged a plan to draw about 7.6 million litres of drinking water a day from a strained aquifer (Municipality of Cerrillos v. Evaluation Commission, 2024). The benefit is global and the water is local, and the people who live beside the turbines are the ones who should decide whether they get built.

Not much alone, and that is no reason to look away. Use the tool when it earns its place and skip it when a template would do. Keep the light model as the default. When your agency buys a tool, put the question into the procurement conversation and ask the vendor for numbers. And if a data center is proposed near the communities you serve, the residents having a say in it is the whole point, not a formality.