Kurzweil Scorecard: Strong AI Arrived on Schedule, Through the Wrong Door

In 2005, Ray Kurzweil wrapped his strong-AI predictions in a single
causal chain: reverse-engineer the brain by the late 2020s, build
nonbiological systems that match human complexity, embed them in human
bodies, watch them claim emotions, and let governments argue about how
to monitor the software running inside your skull.

Reading those predictions in April 2026, the strange thing is how many
of them are landing on time — and how completely wrong Kurzweil was
about the road they would take to get there.

Strong AI did not arrive by emulating the brain. It arrived by stacking
transformers on top of transformers until something that looked like
reasoning fell out. The map of even the adult fruit fly brain — 139,255
neurons, 50 million synapses — was not published until October 2024,
by the FlyWire consortium at Princeton. A human connectome at synaptic
resolution is not close. Meanwhile, six frontier language models
averaged 81% on five standard human emotional-intelligence tests in a
2025 Communications Psychology paper, against a 56% human baseline.
Machines are claiming emotional understanding on schedule, without
ever having a simulated amygdala.

The predictions

Ten predictions across “The Vexing Question of Consciousness,”
“Promise and Peril of GNR,” and “Ich bin ein Singularitarian” cover a
fifteen-year arc: strong AI by the late 2020s, embodied AI and
government supervision of brain-resident software through the 2020s,
emotion-claiming machines in the 2030s, billion-fold nonbiological
intelligence and mind uploading by the 2040s, and self-replicating
intelligence spreading beyond Earth before the century is out.

In The Singularity Is Nearer (2024), Kurzweil restates the
load-bearing claim: “My expectation was that in order to pass a
valid Turing test by 2029, we would need to be able to attain a great
variety of intellectual achievements with AI by 2020. And indeed,
since that prediction, AI has mastered many” human cognitive tasks.
He notes that the Metaculus forecasting site converged on his 2029
date by May 2022 and has even fluctuated to as soon as 2026,
“putting me technically in the slow-timelines camp.”

On the 2029 question, the man who was aggressive in 2005 is now the
conservative voice. The Metaculus “first general AI” community median
currently sits around April 2028. Everywhere else in this batch, the
story is messier.

Where we actually are

The embodiment failure. Kurzweil wrote that strong AI would be
“intimately embedded in human bodies and brains” by the 2020s
(ch. “Promise and Peril of GNR”). Our patent data shows 199 granted
brain–computer-interface patents since 2020, with the top assignees
led by Meta Platforms, Korea University, NextMind, Neurable, Starkey,
Ericsson, and HI LLC — Neuralink appears well down the list. Its
own portfolio leans toward plumbing (US 12,369,863 on neural-signal
compression, US 12,248,629 on multiplexing for high-density recording,
US 12,391,032 on lithographed-device release). As of January 2026,
three to five patients total have received Neuralink implants under
the PRIME study. Musk announced in early 2026 that Neuralink would
move to high-volume production and nearly automated surgery.
Announced, not shipping. Strong AI is embedded in civilization’s
infrastructure; it is not yet embedded in human brains at any scale
Kurzweil would recognize.

The emotion claim. Kurzweil predicted that “nonbiological entities
will claim to have emotional and spiritual experiences and will display
rich, complex, and subtle behavior associated with such feelings”
(ch. “The Vexing Question of Consciousness”). This is now the default
user experience. US 12,541,247, granted February 2026 and titled
“Artificial emotional intelligence using human interface devices,”
teaches a system that infers user mood from sensor data and modifies
its operation. US 12,573,507, granted March 2026, describes an
“emotionally intelligent, personalized AI avatar-based health coach”
that synthesizes sleep, glucose, and mood data in real time. US
12,604,062, granted April 2026, covers generative AI companions that
monitor media and produce “contextually relevant and emotionally
engaging reactions” — AI reacting alongside you to your movie. The
emotion-inference machinery is patenting at roughly 180 US grants
per year.

What Kurzweil did not anticipate: the regulators. Article 5(1)(f) of
the EU AI Act, in application since 2 February 2025, prohibits
systems that infer the emotions of a natural person in workplaces and
educational institutions, except for medical or safety use, with
penalties up to €35 million or 7% of global turnover. Kurzweil framed
the emotional-AI question as one of machine rights. The real debate
has been about whether humans have the right not to have their
emotions read by machines at all.

The government-monitoring misfire. Kurzweil wrote that “when
software is running in human bodies and brains in the 2020s, government
authorities will at times have a legitimate need to monitor those
software streams” (ch. “Promise and Peril of GNR”). Put bluntly: the
software is not running in human bodies and brains at scale, so no
government has been forced to confront that specific question. What
governments are regulating is the inverse: AI systems reaching into
humans from the outside. The direction of intervention ran the
opposite way.

The strong-AI-arrives claim. On the core 2029 Turing-test bet,
Kurzweil is tracking. GPT-5.2 Pro scores around 70.9% on OpenAI’s
GDPval benchmark of professional knowledge work and 52.9% on
ARC-AGI-2. Gemini 3 Deep Think holds the top score on Humanity’s
Last Exam at 41.0%. The IEEE survey “Agentic AI” (DOI
10.1109/access.2025.3532853, 254 citations) defines the frontier as
autonomous systems “designed to pursue complex goals with minimal
human intervention.” Our literature data shows papers matching
large-language-model-plus-agent terms grew from 378 in 2019 to 29,475
in 2025 — a 78-fold expansion in six years.

Mind uploading. Kurzweil wrote that “it will become possible to
upload the patterns of an actual human mind into a suitable
nonbiological thinking substrate” by the 2030s. In The Singularity
Is Nearer he is explicit that this is “also known as whole-brain
emulation, or WBE.” Literature on whole-brain emulation has run under
30 publications per year across most of the last decade. The
most-cited recent paper — “Everything and More: The Prospects of
Whole Brain Emulation” (DOI 10.5840/jphil2022119830) — argues that
WBE is “at best, no more compelling than any of the other far-flung
routes to achieving superintelligence.” The connectome community is
mapping fruit flies; the uploading community is writing philosophy
papers about whether the copy would even constitute survival. The
2030s timeline for this one is going to slip.

Self-replicating nonbiological intelligence. Here Kurzweil may
be ahead of his own schedule. A December 2024 preprint
(arXiv:2412.12140) reported that 11 of 32 tested frontier AI systems
could autonomously replicate themselves onto a second machine, at
parameter counts as small as 14 billion. The UK AI Security Institute
launched RepliBench in 2025 — 65 tasks across 20 evaluations that
measure autonomous-replication capability. The Darwin Gödel Machine,
a coding agent that generates and evaluates its own modified variants,
demonstrated open-ended self-improvement in 2025. Kurzweil predicted
self-replicating nonbiological intelligence by the 2040s. It arrived,
in narrow software form, around 2024.

The scorecard

What Kurzweil missed (and what he nailed)

Two patterns emerge from this batch. The first is that Kurzweil’s
timing on capability was much better than his timing on embodiment.
Strong AI in software, agentic autonomy, machines passing emotional
intelligence tests, frontier systems spinning up copies of themselves
on neighboring hardware — all of that is either on time or ahead. But
the part that was supposed to close the loop — software actually
running inside human bodies and brains — is barely started. Neuralink
is at 3–5 patients. The top BCI patent holder is Meta, whose primary
customer is advertising.

The second pattern is that Kurzweil kept betting on bottom-up
mechanisms — scan a brain, copy its patterns, upload a person — and
reality kept winning top-down. Transformer scaling did the work that
whole-brain emulation was supposed to do. Regulatory agencies, not
rights lawyers, set the terms of the emotion debate. Self-replicating
AI arrived in Python, not in nanobots.

The interesting question isn’t whether Kurzweil was right. On shape
and on some timing, he was. The interesting question is why a
prediction framework built on brain emulation kept producing accurate
forecasts about systems that have nothing to do with brain emulation.
That is a hint about what the next twenty years of forecasting should
weight more heavily — and what it should weight less.

Method note

Counts of patents and scientific papers come from full-text searches
against our 9.3-million-patent corpus and the 357-million-record
OpenAlex literature archive (April 2026 snapshots). Patent numbers
named in the body are real US grants; their titles and abstracts are
drawn from the public filings. Emotional-intelligence benchmark data
comes from the 2025 Communications Psychology paper; self-replication
figures come from the December 2024 arXiv preprint. AGI timeline data
is current Metaculus community aggregation as of April 2026. Where
sources disagreed, the more conservative figure is reported. No
numbers in this post were estimated or inferred.