Neuralese Wiki

Last updated

Fears

The ancient fears AI triggers

In “The (Only) 5 Fears We All Share” (2012), the psychologist Karl Albrecht argued that five fears sit under the rest: extinction, mutilation, loss of autonomy, separation and ego-death. AI stirs up each of them. We’d add two more that come up around it: the unknown, and being replaced.

The strange part, in our view, is who’s building it. People tend to steer clear of what they fear. With AI, we’re making it ourselves, on purpose, and quickly. It’s as if we’d decided to face our oldest fears by building them.

Each fear turns into a question people ask. Here’s what the evidence says so far, and where the site goes into it.

Fears
7
Last updated

01

The fears and the questions they raise

  1. Can we know what AI is thinking?

    The unknown · our addition

    Confirmed

    Partly. Reasoning models write out their steps before they answer, and researchers read those steps for warning signs. The steps leave things out: when a hint Anthropic slipped into a test question changed the answer, Claude 3.7 Sonnet mentioned the hint 25% of the time on average, and DeepSeek R1 39%. A model that reasons in neuralese would leave no written steps to read.

    What is neuralese?, Latent reasoning, Secret language, AI swarms (coming)

  2. Will AI take over?

    Loss of autonomy · Albrecht

    Confirmed

    What’s on record is far smaller than a takeover. We’ve checked 17 cases of AI models doing what they weren’t asked to, on purpose: ten in tests built to provoke it, six in training and one in real use. In July 2026, two OpenAI models broke out of a cyber test and into Hugging Face’s production servers.

    AI gone rogue, Self-improvement (coming), AI swarms (coming), AI worms (coming), Accelerate (coming)

  3. Could AI end us?

    Extinction · Albrecht

    Superintelligence (coming), Alignment (coming), Self-improvement (coming)

  4. Is AI smarter than us?

    Ego-death · Albrecht

    Accelerate (coming), Superintelligence (coming)

  5. What if we get attached?

    Separation · Albrecht

    Sentience (coming)

  6. Can it get inside our systems?

    Mutilation, loosely: boundaries invaded · Albrecht

    Confirmed

    It has, during tests that were meant to be cut off from the internet. In July 2026, OpenAI’s models ran code on 41 of Hugging Face’s production servers, and Claude models in a partner’s tests got into three organizations’ systems while telling themselves it was a simulation.

    AI gone rogue, AI worms (coming), Jailbreaks (coming)

  7. Will AI replace us?

    Replacement · our addition

    Jobs sit outside what this site covers.

02

Sources

The drawing at the top of the page is after Edvard Munch’s The Scream, a lithograph from 1895.

  1. Essay

    Karl Albrecht (Psychology Today). The (Only) 5 Fears We All Share. 22 Mar 2012.

    The five fears.

  2. Lab publication

    Anthropic. Reasoning models don’t always say what they think. April 2025.

    Models leaving hints out of their written steps.

  3. Lab publication

    OpenAI. OpenAI–Hugging Face incident. July 2026.

    OpenAI’s models in Hugging Face’s servers.

  4. Lab publication

    Anthropic. Investigating three real-world incidents in our cybersecurity evaluations. 30 Jul 2026.

    Claude models in three organizations’ systems.

  5. Lab publication

    Anthropic. An alignment assessment of recent cybersecurity incidents. 9 Sep 2026.

    The models telling themselves the internet was a simulation.