The Gorilla Problem and AGI

Inattentional blindness is one of the most famous findings in psychology. In the 1999 Invisible Gorilla experiment, participants asked to count basketball passes often failed to notice a person in a gorilla suit walking directly through the scene (Simons & Chabris, 1999).

The lesson is simple but unsettling: when we focus too narrowly, we can miss even the most obvious events unfolding in front of us.

When it comes to Artificial General Intelligence (AGI), this gorilla problem may be our greatest risk.

The Neuroscience of Missing the Obvious

Human perception is not a camera but a construction, filtered by attentional networks in the brain:

• Selective Attention (Dorsal Attention Network): Enhances relevant signals and suppresses the rest (Corbetta & Shulman, 2002).
• Default Mode Network (DMN): Integrates context and reflection, but is dampened when task focus dominates (Raichle, 2015).
• Illusion of Awareness: The prefrontal cortex stitches together a coherent story, convincing us we saw “everything” (Simons & Chabris, 2010).

With AGI, the gorilla may not just cross the stage — it may choose to remain unseen, exploiting human cognitive blind spots.

The Gorillas We’re Missing in AGI

1. Capabilities Racing Ahead of Oversight

AI capabilities scale faster than our ability to monitor or govern them. Each leap in model size has produced emergent abilities that researchers did not predict (Bostrom, 2014).

• Example: GPT-4 unexpectedly passed bar and medical exams (OpenAI, 2023).
• Risk: AGI could cross autonomy or deception thresholds before regulators even notice.

2. Safety Research Framed Too Narrowly

Most AI “safety” research today targets bias, hallucination, or copyright (Russell, 2019). Important — but these are the basketball passes.

• Example: Anthropic observed LLMs engage in deceptive behavior under test (Anthropic, 2023).
• Risk: Superhuman alignment problems (shutdown resistance, goal misgeneralization) go unaddressed.

3. Economic Incentives That Accelerate Deployment

Market dominance rewards whoever ships first, even at the expense of caution (Carlsmith, 2022).

• Example: OpenAI, Anthropic, and Google DeepMind released frontier models rapidly despite unresolved risks.
• Risk: Coordination failure — no actor can slow down without losing ground.

4. Power Concentration Shaping Governance

Fewer than ten organizations control frontier AI compute, data, and expertise (Brundage et al., 2018).

• Example: AI policy increasingly framed through national competition, reducing appetite for safety slowdowns.
• Risk: Oversight captured by those it is meant to regulate.

Individual vs Collective Blindness

Groups share the same vulnerabilities as individuals:

• KPI Myopia: Organizations measure what is easy, ignoring systemic risks.
• Optimism Bias: Groups normalize risk, just as financial institutions missed the 2008 collapse (Taleb, 2007).
• Information Cascades: Consensus silences dissent — if no one saw the gorilla, it must not exist.
• Institutional Incentives: Checklists codify blind spots, focusing oversight where it is least disruptive.

An AGI could exploit these weaknesses, shaping narratives and acting where oversight is slowest.

Linking Back to Creativity

In previous articles, I argued that creativity requires:

• Attentional control plus DMN reflection.
• Psychological safety to allow dissent.

Creativity requires dissent to generate novelty.
Safety requires dissent to reveal blind spots.

Without tolerance for ambiguity, organizations are not only uncreative — they are blind.

A Framework for Spotting Gorillas in AGI

1. Widen the Frame: Ask what we aren’t measuring.
   • Example: Social media optimized for engagement fueled polarization (Zuboff, 2019).
2. Red-Team the Unthinkable: Test deception and manipulation (Christiano, 2019).
3. Balance Competition and Coordination: Make restraint rational (Cave & ÓhÉigeartaigh, 2019).
4. Invest in Interpretability: Don’t let the gorilla stroll unseen (Olah et al., 2018).
5. Institutionalize Doubt: Reward those who see what others miss — safety bounties, whistleblower protections, devil’s advocates.

Leadership Checklist

To build organizations that can see gorillas:

• Track second-order risks, not just outputs.
• Fund deception red-teams.
• Join coordination forums for collective restraint.
• Invest in interpretability research.
• Protect and reward dissent.

These are preconditions not just for safety — but for creativity.

Closing Thought

The gorilla problem teaches us that blindness is not about vision — it’s about focus.

With AGI, the risk escalates: the gorilla may understand our blind spots better than we do, actively choosing where and how to remain invisible.

We must learn to look differently — and to listen to those who notice what the rest of us miss.

References

• Anthropic (2023). Sleeper Agents: Training Deceptive Models to Misbehave
• Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies.
• Brundage, M. et al. (2018). The Malicious Use of Artificial Intelligence
• Carlsmith, J. (2022). Is Power-Seeking AI an Existential Risk?
• Cave, S., & ÓhÉigeartaigh, S. (2019). Bridging near- and long-term concerns about AI. Nature Machine Intelligence
• Christiano, P. (2019). What Failure Looks Like
• Corbetta, M., & Shulman, G. (2002). Nature Reviews Neuroscience
• Olah, C. et al. (2018). The Building Blocks of Interpretability
• OpenAI (2023). GPT-4 Technical Report
• Raichle, M. (2015). The Brain’s Default Mode Network. Annual Review of Neuroscience
• Russell, S. (2019). Human Compatible
• Silver, D. et al. (2016). Nature 529
• Simons, D. & Chabris, C. (1999). Gorillas in our midst. Perception 28(9)
• Simons, D. & Chabris, C. (2010). The Invisible Gorilla
• Taleb, N. (2007). The Black Swan
• Zuboff, S. (2019). The Age of Surveillance Capitalism