The Signal — August 25, 2026

The most valuable thing a publisher owns right now may be a large pile of text that no machine wrote. Two of today's stories are about what that pile is suddenly worth, and a third is about reading it cheaply.

Thomson Reuters built its own frontier model on Alibaba's open weights

Thomson Reuters announced Thomson at ILTACON on August 24, its first in-house large language model, already running in production inside CoCounsel Legal's Tabular Analysis feature. The company says it spent about $40 million training the model, starting from what the press release calls "a strong open-source foundation" and never names.

CTO Joel Hron named it for Business Insider: the base is a realigned Alibaba Qwen model, arriving by way of an intermediate called Snowdon. He also said the point is to lean less on Anthropic's Claude, which still powers most of CoCounsel. So the shape of the thing is a Toronto legal publisher using Chinese open weights to reduce its dependence on an American lab, which is not how anyone drew the map two years ago.

CEO Steve Hasker says early evaluations put Thomson "on par with the latest frontier models across a range of tasks." That claim is company-authored, measured in-house, and has no third-party confirmation, so treat it as a marketing position rather than a result. The number that actually matters is the $40 million. A July disclosure from the same company put compute spend at roughly $20 million at that point, on less than 10% of its content. Frontier labs spend billions reaching general capability. Thomson Reuters spent the price of a mid-size office building to reach useful capability in one domain, because it already owned Westlaw, Practical Law, Checkpoint and a century of Reuters copy. CoCounsel stays multi-model, with Thomson as the default only where it wins. Watch whether that default spreads.

Sources: Thomson Reuters · Business Insider · SiliconANGLE · Legaltech News · The Next Web

Pew measured how much of the web now reads like a machine wrote it

Pew Research Center's Data Labs published an analysis on August 20 of almost half a million English-language web pages pulled from Common Crawl between 2021 and July 2026, scored with an open detection model called Open Pangram. In the July 2026 sample, 10% of all pages showed significant signs of AI authorship. Narrow that to pages carrying a publication date after ChatGPT shipped and it passes one-third.

The domain split is the part worth keeping. About one in ten .com pages carried those signals, against 4.6% of .org and roughly 1% of both .edu and .gov. In early 2022 all four sat under about 1%. Whatever is filling the commercial web is not filling the university or government web at anything like the same rate.

Pew is unusually careful about what it measured, and the care matters more than the headline. Open Pangram detects linguistic patterns associated with machine writing. It does not establish who wrote anything. And only 10% to 15% of crawled pages carry a machine-readable publication date, so the one-third figure describes dated content, not the web. Both caveats sit in the published methodology, which is more than most detection claims come with. An independent group working from Internet Archive data reached 35% by a different route, which is the kind of agreement that makes a number worth repeating. Set this next to the Thomson Reuters story and the logic gets uncomfortable: as open text fills with machine output, a curated human archive stops being an asset and starts being a moat.

Sources: Pew Research Center · Pew methodology · The Decoder

A small model that stops guessing where entities might be

Fastino released GLiNER2.5 on Monday, and the architecture change underneath it is cleaner than the benchmark table on top. Older extraction models enumerate candidate spans: every possible start, every possible width, scored against every label. Cost climbs with the width grid, so implementations cap how long an entity is allowed to be. GLiNER2.5 scores start and end positions directly instead. The cap disappears, cost scales linearly with document length for a fixed schema, and the model can decode entities and the relations between them in one pass rather than two.

Dropping the explicit span representations also cut memory enough to train on sequences of 4,096 words, which is where the rest of the release comes from: whole contracts and transcripts in a single forward pass, entities of any length, relation extraction that returns a coherent graph, and classification constrained to be schema-valid without a checking layer bolted on afterward. Three checkpoints ship on Hugging Face under Apache 2.0 at 74M, 194M and 287M parameters, all of which run on a CPU.

The scoreboard is less exciting than the design. Fastino reports 56.17 overall macro F1 across 16 zero-shot benchmarks for the multilingual checkpoint, against 56.09 for the previous version. The eight-hundredths of a point separating those is noise. The 24.75-point gain on XNLI is not, and it arrives alongside a slight dip in the same model's extraction average. These are vendor-run evaluations on vendor-chosen benchmarks, so wait for outside replication before believing the ordering. One more detail deserves attention given what Pew just published: Fastino says the training data came from its own generation agent rather than assembled public datasets, on the grounds that no public corpus covers the task and format combinations it wanted. Models trained on machine-written text, to extract meaning from a web increasingly full of machine-written text, is a loop with no obvious exit.

Sources: Fastino · Hugging Face · MarkTechPost

On the Editor's Desk

The Stanford paper on entry-level hiring came close. Brynjolfsson, Chandar and Chen updated "Canaries in the Coal Mine?" with ADP payroll data through June, and the employment gap for 22-to-25-year-olds in AI-exposed occupations widened from 13% to 19%. A wave of coverage this week is what put it in front of us. But the revision is dated August 12, it is still a working paper, and the authors are explicit that the patterns are descriptive rather than causal. A relative gap between occupation groups is not a count of jobs destroyed. We will run it when someone tests the causal claim or the next quarter of data either confirms the widening or does not.

Two others we skipped. The Financial Times has numbers on Anthropic's revenue and on Claude losing ground to cheaper models, and they are genuinely interesting numbers, but the revenue figures come from unnamed sources and the adoption side rests on Ramp billing data, which measures spend among Ramp's own customers rather than the market. We would rather wait for something we can check. And Taiwanese prosecutors indicted nine people over alleged illegal exports of AI servers to China, which matters as enforcement, though an indictment is an allegation and the filings do not say what was shipped or who received it.