Kurzweil Scorecard: The Turing Test and Machine Intelligence

In 2005, Ray Kurzweil bet Mitch Kapor $20,000 that a machine would pass the Turing test by 2029. The bet, administered by the Long Now Foundation, stipulates a two-hour interview — not a casual chat, but a sustained interrogation designed as a proxy for artificial general intelligence. Twenty-one years later, with three years left on the clock, the bet is still unresolved. But the landscape around it has shifted so dramatically that the question is no longer whether machines can fool humans. It’s whether the Turing test still measures anything that matters.

What Kurzweil predicted

The first batch of Kurzweil’s predictions in The Singularity Is Near clusters around a single thesis: machines will reach and then rapidly exceed human-level intelligence, beginning around 2029. In the 2024 follow-up The Singularity Is Nearer, he restated the core claim: “In my 1999 book The Age of Spiritual Machines, I predicted that a Turing test — wherein an AI can communicate by text indistinguishably from a human — would be passed by 2029. I repeated that in 2005’s The Singularity Is Near” (ch. 1). The specific claims range from the near-term — “by 2009, simulated people will still not be up to human standards generally but will be quite satisfactory as sales agents, reservation clerks, and research assistants” (ch. “Deflation … a Bad Thing?”) — to the cosmic: “by the end of the twenty-first century, the nonbiological portion of human intelligence will be trillions of trillions of times more powerful than unaided human intelligence” (ch. “The Six Epochs”).

Between those poles sit a dozen predictions about AI capability timelines, full-immersion VR, brain-computer interfaces, and the eventual merger of biological and nonbiological intelligence. What’s striking, reading them now, is how many of the nearer-term ones are arriving — just not the way Kurzweil described.

Where we actually are

AI benchmarks and the “effective software models” question

Kurzweil predicted that “effective software models of human intelligence will exist by the mid-2020s” (ch. “The Singularity Is Near”). In The Singularity Is Nearer, he doubled down: “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 of humanity’s toughest intellectual challenges.” In April 2026, Claude Opus 4.6 scores 91.3% on GPQA (graduate-level science questions) and 80.8% on SWE-Bench Verified (real-world software engineering). GPT-5, released August 2025, can complete tasks that would take a human programmer three hours. Both models routinely outperform median human performance on standardized tests.

The patent data tracks the buildup. US patent grants mentioning language models, conversational agents, or artificial intelligence benchmarks totaled 117 in the first nine months of 2025 alone — up from a single-digit annual count before 2019. IBM and Microsoft lead with roughly 350 and 235 grants respectively since 2020 in this space, followed by Amazon (115), Google (206 combining entity variants), and Salesforce (60).

In the scientific literature, high-citation papers (20+ citations) on AI benchmarks, the Turing test, and language models exploded from 61 in 2005 to 800 in 2023, with the GPT-4 Technical Report alone accumulating over 2,200 citations since its 2023 release.

By Kurzweil’s own standard — “effective software models of human intelligence” — the mid-2020s prediction looks surprisingly accurate, with one enormous caveat: the mechanism is nothing like what he described. He envisioned reverse-engineering the brain. What happened was transformer architectures trained on internet-scale text. The intelligence is real. The path to it was completely different.

The Turing test and the Long Now bet

The formal Kurzweil-Kapor wager requires a two-hour interview judged by expert panelists — a far harder bar than the five-minute chats that Eugene Goostman “passed” in 2014. As of April 2026, no system has been put through this specific protocol. Prediction markets on Manifold give it roughly even odds of being resolved in Kurzweil’s favor by 2029.

The irony is that the industry has largely moved past the Turing test as a meaningful benchmark. AI companies have stopped reporting MMLU scores because frontier models hit roughly 93%, bumping against the 6.5% error rate in the questions themselves. The benchmarks that matter now — SWE-Bench, GPQA, Terminal-Bench — test whether AI can do useful work, not whether it can fool a human judge in conversation. Kurzweil himself noted that “passing the Turing test will not by itself immediately trigger runaway superintelligence because Turing-level intelligence is closer to an average educated human than to a large team of expert scientists and engineers” (ch. “Runaway AI”). In The Singularity Is Nearer he elaborated: “once an AI does pass this strong version of the Turing test, it actually will have surpassed humans for every cognitive test that can be expressed through language” — but that this would require the AI to “dumb itself down” to avoid revealing its superhuman speed. That distinction is aging well.

The chatbot prediction that arrived late

Kurzweil predicted that “by 2009, simulated people will still not be up to human standards generally but will be quite satisfactory as sales agents, reservation clerks, and research assistants” (ch. “Deflation … a Bad Thing?”). He was about fifteen years early. The intelligent virtual assistant market hit $19.6 billion in 2025 and is projected to reach $99.6 billion by 2031. Customer support chatbots handled 42.4% of the chatbot market in 2024. The AI agent market alone reached $7.6 billion in 2025, growing at 45% annually.

US patent grants for virtual assistants, automated agents, and conversational sales systems grew from 55 per year in 2005 to 400 in 2023. The technology exists and is commercially deployed at scale — just a decade and a half behind Kurzweil’s timeline.

Full-immersion VR: behind schedule, but moving

Kurzweil predicted “full-immersion virtual reality incorporating all the senses will be feasible by the late 2020s” (ch. “Deflation … a Bad Thing?”). In The Singularity Is Nearer, he updated the timeline, noting that “over the next couple of decades, brain-computer interfaces will enable us to experience virtual environments that engage all our senses.” In 2026, haptic suits from bHaptics (40 vibration points) and Teslasuit (electromuscular stimulation) are commercially available but limited to touch and vibration. Smell, taste, and proprioceptive feedback remain experimental. The patent record shows 598 US grants since 2015 mentioning VR with haptic or sensory immersion — real industrial investment, but concentrated in touch and motion rather than full-sensory integration.

This prediction is behind schedule. Haptic feedback is commercial; all-senses immersion is not.

Brain-computer interfaces: the wild card

Kurzweil described “nanobots interacting with biological neurons” that would “create virtual reality from within the nervous system and eventually make virtual reality competitive with real reality in resolution and believability” (ch. “The Singularity Is Near”). In The Singularity Is Nearer, he specified that “we’ll be able to send nanobots into the brain noninvasively through the capillaries.” These nanobot-based neural interfaces remain science fiction. But electrode-based brain-computer interfaces are accelerating fast. Neuralink had implanted 12 patients by late 2025, with users controlling cursors, browsing the web, and playing video games via thought alone. The company plans high-volume production and near-fully-automated surgical implantation in 2026. Its Blindsight implant, aimed at restoring vision, is scheduled for first patient trials this year.

The patent data shows steady growth: 2,120 total US grants for brain-computer or neural interfaces, rising from 17 per year in 2000 to 237 in 2024. Boston Scientific Neuromodulation leads with 32 grants since 2020, followed by Intel (23) and Samsung (27 combining variants).

Kurzweil’s direction was right. His mechanism — nanobots — was wrong. The actual path is electrode arrays and surgical robots.

The scorecard

What Kurzweil missed (and what he nailed)

The pattern across this batch is consistent: Kurzweil’s directional instincts were remarkably good, but his mechanisms were almost entirely wrong. He predicted human-level AI through brain reverse-engineering; it arrived through statistical learning on text. He predicted neural VR through nanobots; the actual frontier is electrode arrays implanted by robotic surgeons. He predicted AI sales agents by 2009; they arrived by 2024 via large language models.

The systematic bias is toward specificity about how and optimism about when. Kurzweil’s exponential models captured the acceleration of capability but couldn’t predict which of many possible technical paths would win. The lesson for technology forecasters: direction is predictable, mechanism is not, and timelines are reliably optimistic by about a decade for consumer deployment — though occasionally dead-on for laboratory milestones.

Three years remain on the Long Now bet. Whether the formal Turing test protocol gets executed before 2029 may matter more than whether the technology is ready.

Method note

This scorecard draws on 9.3 million US patent grants sourced from USPTO bulk grant XML (searched by full-text keyword index), 357 million scientific papers from OpenAlex (filtered for 20+ citations where noted), and current reporting from technology publications accessed via web search in April 2026. Patent assignee counts combine variant spellings of the same organization. AI benchmark figures are from public model evaluations reported by Anthropic, OpenAI, and independent benchmark aggregators. The Kurzweil-Kapor bet status is from the Long Now Foundation’s public ledger. Market size figures are from Mordor Intelligence and industry analyst reports accessed this session.