Mercor’s 29-Benchmark Sample Finds Little Coverage for Tasks Accounting for 90% of U.S. Working Hours

01Mercor’s 29-Benchmark Sample Finds Tasks Accounting for 90% of U.S. Working Hours Have Little or No Coverage

A Mercor study suggests that AI benchmarks still leave much of the American economy untested. Its Benchmark Map finds that tasks accounting for 90% of U.S. working hours receive little or no coverage in an initial sample of 29 benchmarks—not that AI models are unable to perform those tasks.

Mercor built the map to examine whether benchmark scores reflect real work. It used the U.S. Department of Labor’s O*NET taxonomy, which divides the workforce into 923 occupations and 18,797 tasks, then selected 29 benchmarks for their quality and economic relevance.

The company asked a large language model to match each benchmark prompt to any corresponding O*NET tasks. It checked that process against human expert labels for Mercor’s APEX benchmarks, a subset of the 29. A task counted as “covered” only when at least five benchmark prompts were mapped to it.

Mercor then combined those coverage estimates with another study’s model-derived estimates of how much working time Americans spend on each O*NET task. That calculation produced the 90% figure. Within the sample, only 208 of 923 occupations had even one task meeting the five-prompt threshold, while six of 22 occupational families had no qualifying occupation.

Coverage was concentrated rather than random. Tasks rated as exposed to AI were about six times as likely to appear as unexposed tasks. Software engineering and financial analysis were well represented, while truck driving and gas plant operation were sparse. Occupations with larger total wage bills also received more attention, and the most important tasks within an occupation were roughly six to ten times as likely to be represented as the least important.

Even similarly exposed occupations could receive sharply different treatment. Lawyers’ tasks were heavily covered, but tasks performed by legal secretaries and court clerks were barely represented. Mercor said smaller, middle-income knowledge-work occupations appeared particularly undercovered.

The result is provisional. Mercor acknowledged that its 29 benchmarks are neither complete nor perfectly representative, and that an incomplete list necessarily depresses measured coverage by an amount it cannot bound. Its selection criteria include subjective judgments such as “economic realism,” while caps on benchmarks per job family reduce measured coverage in heavily benchmarked fields such as software engineering. Errors outside the APEX subset used to validate the model’s prompt-to-task mappings were not quantified.

Companies and researchers should therefore treat high benchmark scores as evidence about the sampled tasks, not as proof that a model can perform most real jobs. The map measures testing coverage rather than workplace success, and Mercor has not estimated how the 90% figure would change with a more complete benchmark set.

02Current Affairs Puts Ban on AI-Generated Writing and Illustration in Union Contract

Current Affairs, a magazine founded in 2016 that still publishes a print edition, has turned its preference for human-made work into a labor rule. Under a policy adopted as part of its new union contract, the magazine says it may publish neither AI-generated text nor AI-generated illustrations.

The magazine describes the measure as one of publishing’s strictest anti-AI policies. Its stated aim is straightforward: readers should encounter the thoughts of human writers and the work of human artists. Current Affairs says it has released 62 print issues and commissioned hundreds of original illustrations, making creative labor central to both its editorial identity and the physical product it sells.

The restriction is also a response to what the magazine calls the “slopification” of media: replacing writers and artists with cheaply generated material even when quality declines or errors increase. Current Affairs frames that choice as a labor issue because AI can reduce the incentive to pay creators, and as a product-quality issue because its own publishing process depends on work that cannot be reduced to producing copy quickly.

According to the magazine, articles often spend weeks or months moving from draft to publication. Producing the print edition also involves selecting photographs, revising sketches, editing sentences and repeatedly adjusting page layouts. The new rule means published prose must remain human-written, while illustration assignments must continue to result in work made by artists rather than image generators. It also requires editors to scrutinize freelance submissions for possible AI use.

That screening has limits. The editors say they have already received submissions displaying patterns they consider suspicious, but acknowledge that they cannot guarantee every violation will be detected. The available announcement does not provide the contract’s complete language, an enforcement or appeals process, or the consequences for breaking the rule. It also does not say whether the ban covers auxiliary uses such as research, proofreading or transcription when AI-generated material does not appear in the finished publication.

The commitment therefore establishes a clear boundary around what readers see, while leaving some production practices and enforcement details unresolved. Current Affairs has not provided figures showing how the policy will affect costs, staffing or subscriptions after implementation.

03Developer Says AI Comparison of Windows and Linux USB Traffic Produced the Developer’s First Working Linux Scan Within Hours

A developer says an AI-assisted workflow produced a recognizable Linux scan from a Plustek OpticFilm 8200i within hours, after an earlier attempt to write a driver had stalled for months. The scanner’s newer version uses a Genesys GL128 controller, while existing Linux support had targeted an older model carrying the same name but using a GL845 chip.

Writing a driver for undocumented hardware normally requires capturing bus traffic, examining hexadecimal data and testing possible meanings for individual registers or control transfers. The developer’s breakthrough was to give Claude access to a repeatable environment that could run the manufacturer’s Windows driver, record its behavior and compare it with a Linux implementation.

The real scanner was passed through to a Windows virtual machine, but its USB traffic was captured on the Linux host with usbmon. Because usbmon operated below the passthrough layer, the Windows guest could not see the capture process. The developer then performed the same operation using the new Linux driver, recorded it through the same host-side instrument and compared the two traces. Each difference became a candidate driver bug.

According to the developer, the AI could inspect thousands of capture lines, identify repeated sequences and changing bytes, propose what those changes represented, and write tests or driver code to check its hypotheses. The human still chose what to build and verified the results on the physical scanner. The source does not separate how much progress came from the model, the improved tooling or the developer’s own engineering judgment.

The developer’s first working Linux scan arrived within a few hours and was recognizable despite quality problems. Within days, the developer said the driver was nearly complete. It supports color, grayscale, infrared and combined color-infrared scans; resolutions from 300 to 7,200 dpi; 8- or 16-bit channels; infrared-based dust and scratch removal; and an exposure multiplier for dense negatives.

There was a late reversal: SANE, the Linux scanning framework, had independently added support for this scanner in September 2026. In a single-device, single-frame comparison, both implementations resolved the same detail at 3,600 dpi. The developer’s version completed a full-resolution scan in 34.6 seconds versus 54.4 seconds for SANE’s upstream driver, but the author explicitly presented that result as only one data point.

The bespoke scripts were later reorganized into a generic project called hw-harness. No independent reproduction by other developers was reported, leaving its broader usefulness—and any general performance advantage—unproven.