AI lifts the work, not the worker: Google's attorney trial splits its juniors

*A randomized trial of 133 patent lawyers found better output for everyone and no skill gain for juniors, which changes who gets the tool and how you measure it*

Opening read

The most useful evidence today is a three-month randomized trial of patent attorneys from Google Research, and my read is that it undercuts the comfortable story that AI tools level the workforce. Output got better for everyone, senior lawyers got sharper, and junior lawyers split into a stronger group and a weaker one with no average gain. If you are rolling out writing assistants, the measurement you need is the spread of your junior staff's scores, because the average can hide what is happening at either end.

Workforce: the output improved and the junior bench split

The consensus read on AI at work has been that the tools are a leveler, and the common assumption behind it is that the least experienced people have the most to gain. The early evidence that fed that view measured groups, not individuals: across 758 Boston Consulting Group consultants, access to AI lifted performance on routine assignments, and a related BCG and Harvard analysis put the gain at 12% more tasks finished 25% faster when the work sat inside what the model could handle One Useful Thing (https://www.oneusefulthing.org/p/choosing-to-stay-human) Nate Jones (https://www.youtube.com/watch?v=EZ4EjJ0iDDQ). Those are averages for AI users as a group. The leveling story is an inference drawn from them, the guess that the people with the least experience would gain the most, and it is a reasonable one. A smart buyer could conclude that handing every junior a capable assistant is the cheapest training program ever built. That is the inference today's trial tests.

David Autor and Tanya Rodchenko of Google Research ran a randomized controlled trial, meaning lawyers were assigned to the tool by chance so differences can be credited to it, with 133 patent attorneys over three months. The document and the person moved differently. Everyone who used the assistant regularly produced better patent drafts, but only the senior lawyers came out of 90 days with sharper professional judgment than their peers who went without. The juniors are the turn in the story: as a group they ended no more skilled than before, because the spread widened, with more of them scoring high and more scoring low Google Research Blog (https://research.google/blog/does-better-work-always-mean-better-workers/). Better work did not mean better workers across the board.

Here is where I depart from the leveling story. The earlier studies measured the work product, and on that measure the trial agrees: the documents got better. What the trial adds is a separate measure of the person, and on that measure the tool helped the people who already had judgment and divided the people who were still building it. I do not know the mechanism, and the evidence cannot show whether the low tail leaned on the assistant and stopped learning or whether the high tail used it to learn faster. The older consulting data offers a warning sign, though: when consultants hit a problem the model could not handle, those with AI were 19 percentage points less likely to get it right than those working alone One Useful Thing (https://www.oneusefulthing.org/p/choosing-to-stay-human) Nate Jones (https://www.youtube.com/watch?v=EZ4EjJ0iDDQ). That is what over-reliance looks like in a score. Data from more than 400 software organizations points the same way: junior engineers use the tools more often, while staff-level engineers, the most senior hands-on engineers, save about as much time on fewer tokens, the chunks of text a model reads and writes and the unit vendors bill by AI Engineer (https://www.youtube.com/watch?v=Se8jHLliLXE). Experience seems to shape what the tool gives back, and today's evidence sharpens that question without settling it.

Stanford HAI supplies the other half of the picture. Organizations reporting AI use jumped from 55% in 2023 to 88% two years later, so most organizations already report at least some use. The open decision is framing, and Stanford's argument is that staff who are sold generative AI as a way to get more done tend to leave it alone Stanford HAI (https://hai.stanford.edu/news/want-employees-to-embrace-ai-stop-selling-it-as-a-productivity-tool). That connects to the trial in a way the headline numbers miss. A productivity pitch tends to produce disengagement, and an output-only scorecard cannot tell you which juniors disengaged, which leaned on the tool, and which actually learned from it. The framing choice and the measurement choice belong together.

For a business leader deciding who gets the assistant and how success is judged, the stake is that a quality dashboard can show green while part of the junior bench falls behind. My advice is to keep the tool in junior hands but stop treating it as training, and to review the person as well as the document. One concrete check: take whatever your firm already uses to rate the person rather than the work, such as reviewer judgments of each junior's own reasoning or periodic skill assessments, for the junior staff who use an assistant. Instead of comparing the average to the prior period, sort those ratings and compare the top and bottom portions. If the average held while both ends grew, you have the split this trial found, and that is your cue to look at how your juniors are being supported and whether they are learning, not to buy more tooling. If the whole distribution moved up, you have the leveling result and can keep going.

Model pricing: the cuts reach the bottom tier, with a catch in the tiers

The price war that started in late September at the top of the model lineups has now reached the cheapest tier, and the structure of this cut matters more than the headline. The unit is still tokens, the text chunks vendors bill by. Anthropic said on September 28 that Haiku 5.5 would arrive within weeks, and until yesterday its cheapest model was Haiku 4.5 at $1 per million input tokens and $5 per million output tokens Simon Willison's Weblog (https://simonwillison.net/2026/Sep/28/claude-sonnet-5-5/) Simon Willison's Weblog (https://simonwillison.net/2026/Jul/30/luna-price-drop/#atom-everything). The new model resets that baseline, but not evenly. Haiku 5.5 is now out, and Anthropic puts the average saving at roughly 75% against Haiku 4.5, yet the average hides a split: requests up to 100,000 tokens are priced 90% lower, requests above that only 50% lower Anthropic News (https://www.anthropic.com/claude-haiku-5-5). For an engineer deciding what to route to the small model, that means the discount you get depends on how long your requests tend to be, so run the math on your own traffic before you move anything. Short, high-volume calls are the clear win; jobs that feed the model long documents are a much smaller one.

Benchmarks: two gold-level scores, one asterisk, little transfer

NVIDIA's gold-level competition results for Nemotron are real engineering and, on their own, thin buying evidence. A fine-tuned variant of Nemotron-3-Ultra-CC scored 535.4 of 600 on the 2026 International Olympiad in Informatics, the top student programming contest. Fine-tuned means the model was further trained on targeted examples, and this one ran with a custom inference strategy, the method used to steer a model while it answers. That score clears both the gold line of 361.12 and the best human score of 498.27. A Nemotron 3 Ultra setup scored 30 of 42 on the International Mathematical Olympiad, one point above gold, using a generate-verify-refine loop in which the model drafts an answer, checks it, and fixes it. NVIDIA ran the IOI test as it happened, under the same rules the students faced, but that does not make it official: nobody supervised the run, and the score sits outside the contest's real ranking Hugging Face Blog (https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026). That kind of asterisk is not new. Lab-run olympiad claims generally sit outside official judging, and an earlier OpenAI IMO gold claim was made without the official private marking guide Nate Jones (https://www.youtube.com/watch?v=xrzpWXW4-38). For an IT leader weighing an open-weight model for coding, one you can download and run on your own hardware, the catch is what actually scored. The systems were tuned checkpoints, saved versions of a trained model, wrapped in competition-specific scaffolding, the extra software that steers a model through a task. The scores tell you what that setup can do, not what the stock model will do on your work, so they are not enough to buy on by themselves. The more useful signal is the base model's standing on general rankings, where Nemotron 3 Ultra was already rated the strongest U.S. open-weight entrant on the Artificial Analysis Intelligence Index in June OpenRouter Blog (https://openrouter.ai/blog/insights/the-open-weight-models-that-matter-june-2026/). Weight that above the contest scores.

Augur calls

• Call: Google will make Gemini 4 Argon available to paid API customers by November 30, 2026.
• Criteria: An official Google or Google DeepMind announcement or product page lists Gemini 4 Argon as available to paying API customers without a waitlist or partner-only program. Access limited to the Fairwind Program, government users, or an invite list counts as a miss.
• Due: 2026-11-30
• Odds: 55%

What to watch

• Anthropic's tiered Haiku 5.5 pricing will meet real traffic soon, and the thing to track is the 100,000-token break: whether workloads built on long requests see only the smaller 50% cut in practice while short-request traffic gets the full 90%. Anthropic News (https://www.anthropic.com/claude-haiku-5-5)
• Google has kept Gemini 4 Argon inside a limited group of cyber defenders and says it will expand access; the move to paying API customers is the next checkpoint. Nate Jones (https://natesnewsletter.substack.com/p/gemini-4-argon-customer-benchmark)
• Microsoft Research Asia's Agent Lightning v1.0 puts the same agent harness used in deployment, the software wrapper that runs an agent in production, directly inside reinforcement learning training, which is training by trial and reward, in about 3,500 lines of code; follow-on results on production tasks will show whether that approach holds beyond lab benchmarks. Microsoft Research Blog (https://www.microsoft.com/en-us/research/blog/agent-lightning-v1-0-a-3500-line-lightweight-agentic-rl-framework-for-training-agents-with-real-harnesses/)
• OpenAI's global rollout of GPT-6 inside ChatGPT brings faster responses and an Intelligent UI with visuals and interactive experiences; the question for IT leaders is whether company accounts get the same behavior as consumer ones. OpenAI News (https://openai.com/index/gpt-6-for-everyone)
• Ai2's Bolmo work, now published in Nature, comes with new checkpoints that extend its byte-level approach, in which a model reads raw characters instead of word chunks, beyond the Olmo family, and uptake by other model builders would signal whether byte-level models move out of research. Allen Institute for AI (https://allenai.org/blog/bolmo-nature)

Verification notes

The patent attorney findings come from the authors' Google Research blog post; the evidence does not say whether the study is peer-reviewed, and the sample is 133 lawyers in one profession, so the junior split is directional until replicated Google Research Blog (https://research.google/blog/does-better-work-always-mean-better-workers/). Stanford's point about productivity framing is an argument from its article, not a measured result in the evidence. Haiku 5.5's 75% figure is Anthropic's own average. The NVIDIA IOI score is an unofficial run outside the official ranking, and the evidence does not state the IMO run's official status beyond its clearing the official gold threshold Hugging Face Blog (https://huggingface.co/blog/nvidia/nemotron-ioi-and-imo-2026).

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-10-08?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-10-08)