Independent cognition research · est. 2019

Can Machines Think Like Humans?

Computation is settled. Understanding is not. We study the gap between outputs that look intelligent and systems that actually reason — and turn the evidence into decisions you can defend.

  • 34 peer-reviewed studies mapped
  • 120+ benchmark tasks
  • No vendor funding
Scroll

About the inquiry

The question is not whether machines compute. It is whether they understand.

A model can finish your sentence without knowing what a sentence is. That gap — between fluent output and grounded comprehension — is where the real argument lives.

Since Turing proposed his imitation game in 1950, the debate has moved through symbolic reasoning, connectionism, and now large-scale statistical learning. Each wave produced systems that behave intelligently in narrow conditions and fail, often revealingly, just outside them.

We do not argue from philosophy alone. Our work pairs classical thought experiments — the Chinese Room, the symbol grounding problem, the frame problem — with measurable behaviour: compositional generalisation, causal intervention, calibration under distribution shift, and the capacity to represent uncertainty.

Our position is deliberately modest. Today's systems exhibit functional reasoning in bounded domains, and the burden of proof for genuine understanding remains unmet. What matters is not settling the label, but knowing precisely where the boundary sits for your use case.

  • Functionalism
  • Biological naturalism
  • Predictive processing
  • Enactivism
  • Symbol grounding
AI image prompt: a quiet reasoning laboratory at night, floating holographic logic diagrams, chalkboard equations, brushed steel and warm amber desk lamp, deep shadows, wide-angle editorial photography, no text, no people, no women.
Fig. 02 — Protocols before opinions: the room where every claim is tested twice

What we do

Evidence for teams deciding what to believe about AI.

Six focused engagements. Each one replaces a vague claim — “it reasons”, “it hallucinates”, “it is safe” — with a testable one.

01 / 06

Cognitive architecture review

We map how your system represents goals, memory, and uncertainty, then identify which failures are architectural rather than incidental.

  • Audit
  • Systems
  • 6 weeks

02 / 06

Benchmark design & red-teaming

Existing suites reward fluency. We build adversarial tasks that isolate reasoning from pattern completion, and report what survives.

  • Evaluation
  • Adversarial
  • 4 weeks

03 / 06

Human–machine comparative studies

Matched protocols for people and models, so you can say where performance is comparable, where it diverges, and why the difference matters.

  • Behavioural
  • Baselines
  • 8 weeks

04 / 06

Interpretability sprints

Probing, activation analysis, and causal tracing on the specific behaviours your product depends on — not generic explainability theatre.

  • Mechanistic
  • Probing
  • 3 weeks

05 / 06

Reasoning & alignment audits

Independent review of chain-of-thought reliability, sycophancy, specification gaming, and the gap between stated and revealed objectives.

  • Safety
  • Governance
  • 5 weeks

06 / 06

Executive briefings & workshops

A grounded, jargon-free account of what current systems can and cannot do, tailored to the decisions your leadership actually faces.

  • Advisory
  • Training
  • 1–3 days
AI image prompt: a translucent synthetic cortex model on a dark laboratory plinth, layered glass and copper circuitry, soft violet and cyan rim lighting, macro photography, dust particles, ultra-detailed, no text, no people, no women.
Fig. 03 — A model of a model: structure without a subject

We test the claim, not the demo.

Every engagement begins with a falsifiable question and ends with a document your engineers and your board can both use. No black boxes explaining black boxes.

  • 01Pre-registered protocols and held-out task families.
  • 02Human baselines collected under identical constraints.
  • 03Uncertainty reported, not hidden behind a single score.
  • 04Reproducible artefacts, seeds, and prompts included.

Client reviews

What it is like to work with us.

Six engagements, six honest assessments. Ratings are collected after delivery, not during the pitch.

Statistics

The evidence, quantified.

Aggregated results from our 2024–2026 reasoning programme. Every figure is traceable to a published protocol.

0+

Benchmark suites evaluated

0%

Inter-rater agreement

0

Research partners

0k

Reasoning traces analysed

Comparative results

Human vs. model reasoning

Human baseline Best model

Accuracy (%), mean of three seeds, temperature 0.2. Model results use the strongest publicly available checkpoint in each family at time of testing.

Scaling behaviour

Capability vs. compute

Aggregate score across 42 reasoning tasks, plotted against log training compute. Gains flatten sharply after the final decade of scale.

AI image prompt: abstract visualization of scaling laws, luminous gradient mountain ridges made of data points rising into a starless indigo sky, thin cyan contour lines, minimal cinematic composition, ultra-detailed, no text, no people, no women.
Fig. 04 — Emergence is a shape, not a promise

Legal

Privacy Policy

Legal

Disclaimer

Contact

Bring us a question worth testing.

Tell us what you are trying to decide. We reply to substantive enquiries within two business days.

You do not need a finished hypothesis. A suspicion, a failed evaluation, or an uncomfortable board question is enough to start.

We work with research teams, product groups, regulators, and leadership teams who need an independent read on machine reasoning — before a launch, a procurement, or a public claim.

  • Scope calls are 30 minutes and carry no obligation.
  • We decline engagements where the evidence cannot be published in some form.
  • No sales sequences, no newsletter enrolment, no follow-up automation.

Unsubscribe

Leave the signal.

One step, no questions. Enter the address you subscribed with and we will remove it from every list we control.

Removal is processed immediately. Transactional messages you have requested separately are not affected.