iiinova — where your digital self keeps evolving

Reposted by Eyas Ayesh

Anna (Anya) Ivanova @neuranna.bsky.social · 10/08/2026

Our #CogSci2026 paper on **self-reported variations in inner language** got the Disciplinary Diversity & Integration Award (philosophy + cogsci!) Paper: escholarship.org/uc/item/1qr6... +a journal-length preprint coming soon By Raquel Krempel, Jin Li @ruimingao.bsky.social @ruifengyu.bsky.social

escholarship.org

2245

Reposted by Eyas Ayesh

neurotaha @neurotaha.bsky.social · 03/08/2026

Excited to be at #CCN2026 in NYC! I’ll be presenting our poster, “Divergent Scaling Patterns in the Auditory Cortex vs. Other Brain Regions Reveal Distinct Sources of LLM–Brain Alignment,” with @neuranna.bsky.social 📌 Poster B78 · Session B 🗓️ Tuesday, August 4 Come say hi and check it out!

083

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

We conclude with a call to action: AI deception research should be mechanism-sensitive. Rather than seek catch-all solutions, researchers should identify the capabilities and representations driving deception and use them to develop targeted monitoring, intervention, and mitigation strategies.

100

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

Box 2: Understanding a model’s cognitive-capabilities can inform us about its capacity for deception (can it deceive?), but not its propensity (will it?). The latter requires us to study agentic environments and how they incentivize deception. We briefly explore some of these incentives.

100

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

Two additional points. Box 1: Without mechanistic evidence, classifying deceptive behavior may be difficult. Evaluators can use either a parsimonious heuristic, assigning the lowest plausible level, or a precautionary one, assigning the highest, depending on their goals.

100

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

Language Fluency: the ability to generate plausible, strategically framed communication. Linguistic deception extends beyond just producing believable lies. An agent can hypothetically deceive via omission, implication, ambiguity, framing, or technically true but misleading statements.

100

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

World Modeling: the ability to model and simulate the environment. Understanding the current state of the world, simulating plausible alternatives, and predicting how different actions would affect future outcomes can greatly enhance an agent's ability to successfully deceive.

100

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

Theory of Mind: the ability to model another agent’s beliefs, goals, and internal mental states. Explicit deception requires some ability to reason about what another agent currently believes, what belief the deceiver wants them to adopt, and how a particular action might change current beliefs.

100

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

Intentionality: an agent’s capacity for deliberate action. We break this down further into the ability to represent a goal, plan how to achieve it, and selectively execute or suppress actions in service of that goal.

110

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

We identify four core competencies that support explicit deception: intentionality, theory of mind, world modeling, and language fluency.

110

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

Then, we focus on the most sophisticated level—explicit deception—where an agent acts with the deliberate goal of inducing a false belief in another agent. We ask: what cognitive capabilities are necessary for, or would greatly enhance, explicit deception?

100

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

First, we extend Mitchell’s 1986 framework for characterizing animal deception to AI systems, drawing parallels between animal and AI deception at each level. This framework classifies deceptive actions by their cognitive complexity, ranging from automatic or preprogrammed to deliberate.

110

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

In our new preprint, we take a cognitive approach, drawing on deception research in animal cognition and human cognitive psychology to identify cognitive capabilities that enable deception in AI systems.

100

Eyas Ayesh @eyasayesh.bsky.social · 20/07/2026

🚨New position piece! “Deception-Enabling Cognitive Capabilities in AI Systems” 🧠🤖 w/ @dobyrahnev.bsky.social @apurvaratan.bsky.social @neuranna.bsky.social To better identify and mitigate deceptive AI behaviors, we need a deeper understanding of their complex and multifaceted nature...🧵👇

1115