Start free trial

Five minutes.
AI, made clearer.

An asylum tool checked on 0.3% of its work, a #1 that leads four of nine tests, and a rebuild that fails 81.5% of the time.

Read the briefing

Members read every story in the archive, re-verified. Sources and corrections stay open to everyone.

What the pen marks mean
  • rememberThe takeaway: what to remember
  • figureThe number that matters
  • evidenceThe evidence line
  • rulingThe ruling
  • claimThe claim that fails

Each mark is drawn once, as you reach it. Numbers are only circled when they come from the story’s own figures.

The Home Office's asylum AI failed 45 of 203 expert checks after rollout. Experts checked about 0.3% of the summaries.

A lone inspector lifts one sheet with tweezers from a flood of summaries rushing past on a conveyor belt.
AI illustration

Open Rights Group published Home Office figures showing 45 of 203 sampled AI interview summaries failed the first-stage expert check after the tool was rolled out in April 2026.

Members · 30 days free

See what the evidence actually shows.

Members read the full check on every story: what the evidence shows, why it matters to you and the one line to take with you. Every past edition, re-verified, and the Receipts Pack come with it. A$89 a year, about A$0.24 a day.

  1. At sign-up: 30 days free. Payment card required. One introductory trial per customer.
  2. Your trial ends 30 days after you start. Unless you cancel before then in Account → Manage subscription, we charge A$89 for the first year. It renews automatically at A$89/yr until cancelled.
  3. You can ask for a full refund within 14 days after any annual payment, renewals included, with no reason needed, by emailing [email protected]. This voluntary refund does not limit your rights under the Australian Consumer Law.

By starting your trial, you agree to the Terms.

Inferno Tech Pty Ltd t/a BS Killer, ABN 27 647 413 474

Open the evidence4 source pages

The claim we checked

AI summaries of asylum interviews save Home Office decision-makers 23 minutes per case, with only a small proportion (9%) of summaries found inaccurate in the pilot.

These are quoted receipts, not a count of independent investigations. Several reports may rely on the same original source.

gov.uk ↗
the pilots suggest that the ACS tool could save 23 minutes per case when reviewing interview transcripts
gov.uk ↗
A small proportion of summaries produced (9%) were deemed to be inaccurate or had missing information
openrightsgroup.org ↗
In a first-stage SME check on 203 summaries, 139 passed, 19 were flagged for further review, and 45 failed (22%).
openrightsgroup.org ↗
Just 0.3% of total generated summaries between April-July were checked by Home Office Subject Matter Experts (SME).
theregister.com ↗
summaries could pass the checks even if they omitted important details, as long as the information they did include was accurate.
Open this check in the full collection →
The idea, illustrated01

Go inside the check.

    The full explanation appears when your reading access is confirmed.
    Next: Gemini 4 Carbon's #1

    Gemini 4 Carbon is #1 on Artificial Analysis, and the tester agrees. It leads 4 of the 9 tests, and most people cannot use it yet.

    A runner holds a giant trophy on a podium in front of a partly lit scoreboard while the stadium gates stay chained to the waiting crowd.
    AI illustration

    Google launched Gemini 4 Carbon and said it ranks #1 on Artificial Analysis Intelligence Index v5 at 59.

    Members · 30 days free

    See what the evidence actually shows.

    Members read the full check on every story: what the evidence shows, why it matters to you and the one line to take with you. Every past edition, re-verified, and the Receipts Pack come with it. A$89 a year, about A$0.24 a day.

    1. At sign-up: 30 days free. Payment card required. One introductory trial per customer.
    2. Your trial ends 30 days after you start. Unless you cancel before then in Account → Manage subscription, we charge A$89 for the first year. It renews automatically at A$89/yr until cancelled.
    3. You can ask for a full refund within 14 days after any annual payment, renewals included, with no reason needed, by emailing [email protected]. This voluntary refund does not limit your rights under the Australian Consumer Law.

    By starting your trial, you agree to the Terms.

    Inferno Tech Pty Ltd t/a BS Killer, ABN 27 647 413 474

    Open the evidence4 source pages

    The claim we checked

    Google's Gemini 4 Carbon ranks #1 on Artificial Analysis Intelligence Index v5, making it the best AI model available.

    These are quoted receipts, not a count of independent investigations. Several reports may rely on the same original source.

    blog.google ↗
    Carbon ranks #1 on the recently updated v5.0 of Artificial Analysis Intelligence Index at 59.
    artificialanalysis.ai ↗
    Gemini 4 Carbon leads the Intelligence Index v5, with the #1 score in 4 of the 9 Index evaluations
    artificialanalysis.ai ↗
    Carbon makes substantial gains elsewhere but is not leading in coding and knowledge work
    blog.google ↗
    Carbon jumps +89 elo points to land at #4 (1768) on WebDev Arena
    blog.google ↗
    Starting this week, we will make Carbon available to a subset of cyber defenders
    artificialanalysis.ai ↗
    Gemini 4 Carbon is being rolled out to selected users and is not yet publicly available.
    blog.google ↗
    at $2 per million input tokens and $10 per million output tokens
    blog.google ↗
    After the introductory period expires, the price of $4 per 1M input tokens and $20 per 1M output tokens will apply.
    artificialanalysis.ai ↗
    All our charts and figures use standard pricing for both Gemini 4 Carbon and Argon.
    artificialanalysis.ai ↗
    On average, it costs $7.76 per task to evaluate Gemini 4 Carbon (High) on the Intelligence Index.
    artificialanalysis.ai ↗
    somewhat expensive when comparing to other models of similar price. It's also faster than average, however very verbose.
    news.ycombinator.com ↗
    Gemini 4 Carbon | Hacker News
    Open this check in the full collection →
    The idea, illustrated02

    Go inside the check.

      The full explanation appears when your reading access is confirmed.
      Next: Opus 5.5 on ProgramBench

      Claude Opus 5.5 hit 18.5% on ProgramBench, up from zero in May. It cloned 37 programs it could run and probe.

      A robot carpenter builds a copy of a working music box beside the original while an inspector checks only a few keys.
      AI illustration

      A trending post said Opus 5.5 can rebuild real programs from scratch, citing 18.5% on ProgramBench after 0% in May.

      Members · 30 days free

      See what the evidence actually shows.

      Members read the full check on every story: what the evidence shows, why it matters to you and the one line to take with you. Every past edition, re-verified, and the Receipts Pack come with it. A$89 a year, about A$0.24 a day.

      1. At sign-up: 30 days free. Payment card required. One introductory trial per customer.
      2. Your trial ends 30 days after you start. Unless you cancel before then in Account → Manage subscription, we charge A$89 for the first year. It renews automatically at A$89/yr until cancelled.
      3. You can ask for a full refund within 14 days after any annual payment, renewals included, with no reason needed, by emailing [email protected]. This voluntary refund does not limit your rights under the Australian Consumer Law.

      By starting your trial, you agree to the Terms.

      Inferno Tech Pty Ltd t/a BS Killer, ABN 27 647 413 474

      Open the evidence3 source pages

      The claim we checked

      Claude Opus 5.5 can now rebuild real programs from scratch: 18.5% on ProgramBench, up from 0% in May.

      These are quoted receipts, not a count of independent investigations. Several reports may rely on the same original source.

      vals.ai ↗
      Claude Opus 5.5 ranks first on ProgramBench with 18.50%
      programbench.com ↗
      Given only a compiled binary and its documentation, agents must architect and implement a complete codebase that reproduces the original program's behavior.
      arxiv.org ↗
      We evaluate 9 LMs and find that none fully resolve any task
      vals.ai ↗
      a task is counted as Fully Resolved only when the submitted implementation passes all tests.
      vals.ai ↗
      Claude Opus 5.5 87.0% 130 37
      programbench.com ↗
      The agent can run the given program with any input and observe exactly what it does
      programbench.com ↗
      Note that behavioral tests can never cover all possible inputs.
      programbench.com ↗
      ranging from small terminal utilities like jq and ripgrep to massive software projects like the PHP compiler
      arxiv.org ↗
      Models favor monolithic, single-file implementations that diverge sharply from human-written code.
      Open this check in the full collection →
      The idea, illustrated03

      Go inside the check.

        The full explanation appears when your reading access is confirmed.
        The finish line

        Edition complete

        You’re up to speed.

        That’s the 29 Oct 2026 briefing. Keep the useful bits. Leave the noise.

        Reading estimate: 511 words at 200 words per minute. Source quotes and the optional sections below add reading time.

        Get tomorrow's check in your inbox.

        One AI claim a night, checked against independent sources, receipts attached. Free.

        Free email updates. Unsubscribe any time.