PulseAugur
EN
LIVE 21:53:52

AI Audits Must Be Sealed at Execution, Not Annual Checks · 5 sources tracked

A series of posts on Mastodon argues for a shift in how AI models are audited and verified. The author contends that traditional annual audits are insufficient due to model drift and retraining, advocating instead for cryptographic provenance and sealed records of AI decision-making processes. This approach, particularly using post-quantum cryptography, is presented as essential for proving model actions and is becoming a critical factor in areas like insurance pricing. AI

IMPACT Proposes a new standard for AI verification, potentially impacting how AI systems are insured and regulated.

RANK_REASON The cluster consists of opinion pieces discussing AI auditing and verification methods, rather than a specific release or event.

Read on Mastodon — mastodon.social →

AI-generated summary · Google Gemini · from 5 sources. How we write summaries →

AI Audits Must Be Sealed at Execution, Not Annual Checks · 5 sources tracked

COVERAGE [5]

  1. Mastodon — mastodon.social TIER_1 English(EN) · Mickai ·

    Continuous Audit or No Audit at All I think the annual audit is already a fiction. Models drift and get retrained between visits, so a yearly stamp certifies a

    Continuous Audit or No Audit at All I think the annual audit is already a fiction. Models drift and get retrained between visits, so a yearly stamp certifies a system that no longer exists. I argue assurance must be sealed at execution, or it is nothing. https:// mickai.co.uk/art…

  2. Mastodon — mastodon.social TIER_1 English(EN) · Mickai ·

    The Validator Economy Is a Verification Market, Not a Mining Race We have been pricing the wrong thing. I want validators paid not to order transactions but to

    The Validator Economy Is a Verification Market, Not a Mining Race We have been pricing the wrong thing. I want validators paid not to order transactions but to prove a machine acted within its authority. That is the market I am building, and here is why it changes who gets paid. …

  3. Mastodon — mastodon.social TIER_1 English(EN) · Mickai ·

    Adversary Models in Friendly Clothing I have learned to distrust a model by how friendly it looks. The handle fits your grip while the steel traces back somewhe

    Adversary Models in Friendly Clothing I have learned to distrust a model by how friendly it looks. The handle fits your grip while the steel traces back somewhere else. I argue cryptographic provenance over the whole lineage is the only border that holds. https:// mickai.co.uk/ar…

  4. Mastodon — mastodon.social TIER_1 English(EN) · Mickai ·

    When the Inference Itself Must Carry a Signature I think most signed-AI is signing the wrong thing. We seal the words and leave the thinking deniable. My argume

    When the Inference Itself Must Carry a Signature I think most signed-AI is signing the wrong thing. We seal the words and leave the thinking deniable. My argument: bind the whole decision event, model, inputs and authority, under post-quantum strength, or you have proven nothing.…

  5. Mastodon — mastodon.social TIER_1 English(EN) · Mickai ·

    Your Insurer Will Ask to See the Record I have watched AI cover get repriced on a single question: can you prove what the model did. My view is that a sealed, p

    Your Insurer Will Ask to See the Record I have watched AI cover get repriced on a single question: can you prove what the model did. My view is that a sealed, post-quantum audit record is now an actuarial input, not a compliance nicety. https:// mickai.co.uk/articles/your-ins ure…