“If you’re going to make the internet comparison, it’s like we’re in 1997. It’s very exciting. Most stuff kind of doesn’t work yet.” – Benedict Evans – Independent analyst

The recurring difficulty in technology cycles is distinguishing genuine platform shifts from speculative hype before the evidence is obvious. In artificial intelligence, the tension is acute: spectacular demonstrations sit alongside brittle systems, unclear business models and unpredictable cost curves, leaving practitioners unsure whether to treat current systems as infrastructure or experiments.^1,3 The comparison with the commercial internet in the late 1990s is not about nostalgia; it is about how industries mis-price uncertainty when a technology is clearly powerful but operationally incomplete.

Early platform moments and radical uncertainty

The late 1990s were a period when the web was clearly going to be important, but almost no-one could specify how, for whom, or on what economic terms.³ Large incumbents funded speculative online divisions; start-ups raised capital on traffic metrics; and regulators struggled to map law designed for physical trade onto digital flows. Much of the infrastructure that would eventually make the internet ubiquitous either did not yet exist or was only partially deployed: broadband penetration was low, mobile data was negligible, and payment rails were fragmented.

The current AI cycle exhibits a similar structure of known significance, unknown specifics.^1,3,9 On any reasonable reading of the research trajectory, the capability frontier in machine learning has shifted substantially: models that perform language understanding, image generation, code synthesis and multi-modal reasoning at usable levels were simply not available a decade ago.⁴ Yet across most domains outside software development, robust product-market fit is partial at best.⁵ Organisations sense that ignoring AI is strategically dangerous, but cannot reliably determine which use cases will endure or how value will be captured.

This pattern produces what Benedict Evans describes as a posture of “radical uncertainty”: it is rational to assume that AI is transformative on the scale of the internet or mobile, while simultaneously accepting that most current artefacts are prototypes rather than settled products.^3,7,9 The strategic challenge is therefore allocating capital, attention and organisational change in a context where timelines, winners and margins are all contested.

Why “most stuff kind of doesn’t work yet”

Describing AI systems as “not working” is not a claim that they fail in all tasks; it is a statement about their reliability, scope and integration. In 1997, web pages loaded slowly, browsers crashed, payment systems were clunky, and basic actions like finding information or completing a purchase were error-prone compared with offline alternatives. Nevertheless, those fragile services pointed to entirely new ways of doing media, retail and communication.

Current AI systems exhibit analogous fragility. Large language models hallucinate facts, struggle with long-horizon reasoning, and degrade under distribution shift; generative image systems mis-handle edge cases and embed training-data biases; and applied models often require extensive prompt engineering or guardrails to behave within acceptable safety boundaries.^2,4 For many enterprise workflows, this yields value only when a human is firmly in the loop, constraining automation and complicating ROI calculations.

The problem is compounded by infrastructure maturity. In many organisations, data is siloed, poorly labelled, or inconsistent, limiting the performance of domain-specific models and raising governance risks. Operationalising AI therefore demands simultaneous progress in data architecture, security, compliance and change management, not just model capability. The result is a landscape where demonstration projects look impressive, but production deployments remain narrow and fragile.

Scale of impact: “as big as the internet or mobile”

One of Evans’ more controversial positions is that AI is “as big as the internet or mobile, and only as big”.^3,7,9 That formulation rejects both minimisation (“just another feature”) and existential inflation (“beyond historical comparison”). By anchoring AI against the observed effects of the web and smartphones, the argument focuses analysis on what platform-scale change empirically looks like.

The internet restructured distribution, lowered search costs and made information abundant. Mobile compressed those effects into personal context, creating continuous connectivity and location-aware services. AI, in this framing, is a third layer: pervasively embedding statistical inference and pattern recognition into interfaces, processes and decisions.³ Instead of treating models purely as centralised enterprise tools, the trajectory points towards AI-level capabilities being baked into most digital experiences, much as networking and touch screens are today.

Thinking in platform terms matters because it changes strategic questions. The relevant issues are not whether a particular chatbot will dominate, but how AI affects cost structures, organisational design, regulatory regimes and competitive moats. As with the early internet, most enduring value is likely to be captured not by the first visible applications, but by businesses that correctly infer how the underlying capabilities alter industry economics.

The 1997 internet analogy unpacked

The comparison to 1997 does specific analytical work. That year sits after the web ceased to be fringe, but before search, social media, broadband and smartphones resolved the consumer experience into a mature pattern. For AI, the equivalent moment is one where models function well enough to be widely usable, but the larger stack – developer tools, standards, organisational practices, and complementary technologies – is not yet aligned.^3,9

Several features of 1997 are salient:

• Infrastructure present but partial. Core protocols existed and browsers were mainstream, but bandwidth, devices and hosting were constrained. Today’s AI picture is parallel: transformer architectures, large-scale training and inference APIs are broadly available, yet context windows, latency, cost and tooling still limit what can be built.
• Business models speculative. Early web firms monetised via banner advertising, subscription experiments or pure growth narratives. Current AI ventures are similarly divided between usage-based pricing, bundled SaaS add-ons, and pure platform plays, with unit economics often contingent on future cost curves.
• Regulation and norms unsettled. Data privacy, jurisdiction, content liability and taxation for online activities were unclear in the 1990s. Today, policymakers struggle with AI-related copyright, safety standards, labour impacts and competition law, producing uncertainty for investors and operators.

Positioning AI in “1997” therefore emphasises that a long, messy phase of experimentation and infrastructure building lies ahead. For participants, the implication is not to wait for clarity, but to recognise that present conditions are inherently provisional.

Jobs, automation and misplaced apocalypses

The analogy also reframes anxiety about employment. Each major technology cycle generates predictions of wholesale job destruction, often extrapolating from visible capability gains without accounting for institutional adaptation. Evans argues that fears of an imminent “job apocalypse” from AI repeat patterns seen around the internet and automation: genuine disruption, but not instantaneous collapse.^3,9

The early web did eliminate certain roles – for instance, aspects of travel agency work, classified advertising or manual back-office processes – yet it also created new occupations in web design, digital marketing, e-commerce operations and software engineering that were hard to foresee ex ante. Similarly, mobile produced app ecosystems, gig work structures and location-based services. AI is likely to track this mixed pattern, altering task composition within roles as much as entire job categories.

Critically, the current evidence suggests that coding is the clearest domain with strong product-market fit: developers using AI tools report substantial productivity improvements, even if these are not yet perfectly measured.⁵ Other domains, such as legal drafting, medical documentation or customer service, exhibit promising pilots but face heavier constraints from regulation, liability and organisational inertia. This divergence reinforces the idea that broad job impacts will be staggered and sector-specific, rather than uniform and immediate.

Where value may accrue

A central strategic question is who ultimately captures the surplus from AI deployment. During the internet’s maturation, value pooled around a few horizontal platforms (search engines, social networks, cloud providers) and a series of vertical category leaders in e-commerce, media and software. Many early firms failed, not because demand for internet services vanished, but because they misjudged timing, economics or defensibility.

In AI, there is an emerging stack with at least three economic layers:

• Foundation and infrastructure: model providers, training hardware, data centres and orchestration tooling.
• Horizontal application platforms: productivity suites, developer tools and integration frameworks embedding AI into general workflows.
• Vertical and niche applications: sector-specific products built on top of models and infrastructure.

The 1997 framing suggests that dominant players in each layer may not yet exist, and that switching costs, standards and regulatory constraints will significantly shape outcomes. For organisations, the safer assumption is that AI capability becomes ubiquitous and commoditised at the infrastructure level over time, shifting differentiating power towards proprietary data, domain knowledge and distribution.

Technical and epistemological challenges

One under-discussed aspect of the “most stuff doesn’t work yet” observation is the epistemic instability of current AI techniques. Machine learning systems are fundamentally empirical: they learn statistical relationships from data rather than executing explicitly coded rules. A recent analysis argues that AI inherits a longstanding crisis from psychology, driven by the lack of a unified object of study, fragmented tasks and methodological eclecticism.²

This creates several practical difficulties. First, many models lack transparent mechanisms for explaining their decisions, complicating deployment in domains requiring accountability. Secondly, success in benchmark tasks does not necessarily translate into robust performance under real-world conditions, particularly when human behaviour responds strategically to the presence of AI systems. Thirdly, the relative success of existing approaches can discourage deeper critical reflection on their limitations.²

These challenges are not fatal, but they imply that reliability and trust will remain contested for some time. As with security and privacy on the early web, repeated failures are likely to drive iterative improvements in architecture, tooling and governance. However, the process will be uneven, and some classes of application may prove far harder to stabilise than optimistic early demos suggest.

Analogy as a tool for strategic reasoning

The use of historical analogy is not simply rhetorical. Cognitive science research indicates that analogical reasoning is central to human problem-solving, allowing us to project structure from familiar domains onto unfamiliar ones.^6,14,17 When analysts invoke the 1997 internet to describe AI, they are engaging in a specific form of analogical mapping: identifying relational similarities (early infrastructure, speculative business models, regulatory lag) while acknowledging differences in detail.

Work on analogy shows that the level of abstraction at which similarities are represented strongly affects how useful the analogy becomes.^12,17 A superficial comparison (e.g. “lots of start-ups and hype”) yields little guidance. A relational comparison – focusing on mechanisms such as network effects, cost declines, and complement development – can support more disciplined scenario planning. In this context, the 1997 analogy helps strategists think about:

• How long it may take for AI to become boring infrastructure embedded in everything.
• Which institutional adaptations (regulatory, organisational, educational) are typically required for a technology to stabilise.
• Where mispricings of risk and opportunity are likely to occur, based on past cycles.

Of course, analogies can mislead. The internet did not involve models making probabilistic inferences on human language at scale, nor did it raise identical safety questions about system autonomy or alignment. Analysts must therefore use historical comparison as a starting point, not an endpoint, testing where structural differences break the mapping.

Why the moment is strategically uncomfortable

For decision-makers, the present AI landscape is uncomfortable precisely because the technology is both too promising to ignore and too immature to plan around with confidence. Capital markets, media and internal stakeholders demand clear strategies and timelines, yet the rational stance is that many current assumptions – about costs, architectural patterns, dominant vendors and regulatory regimes – may be wrong.

Evans’ suggestion that people stop hiding from the technology and instead start using it reflects a pragmatic response to this discomfort.^3,7,9 In an environment of radical uncertainty, learning-by-doing becomes a critical hedge: small-scale experimentation, capability building and cultural acclimatisation improve an organisation’s option value irrespective of which specific AI paradigm wins. This echoes the late-1990s behaviour of firms that invested early in web literacy, internal tooling and online branding, positioning themselves to move faster once the infrastructure matured.

Debates and objections

The 1997 comparison is not universally accepted. Critics argue that AI may be more discontinuous than the internet, with potential to disrupt cognitive labour in ways that lack historical precedent. Others suggest that the pace of diffusion is faster today, given cloud infrastructure and existing digital workflows, implying that the “early” phase will be shorter. There are also concerns that safety and alignment issues make AI qualitatively different from previous platform shifts, requiring more conservative deployment.

Defenders of the analogy respond that earlier technologies also provoked existential anxieties, from industrial automation to nuclear power, and that the eventual pattern was mixed: significant disruption, but embedded within broader institutional evolution.¹⁶ They contend that viewing AI as another major, but not unbounded, platform helps avoid both complacency and catastrophism, enabling more tractable discussions about governance, regulation and economic impact.

Ultimately, the usefulness of the 1997 framing depends on how seriously one takes radical uncertainty. If the future path of AI is assumed to be largely determined and visible, the analogy looks unnecessarily cautious. If, however, one accepts that current systems may be several iterations away from their mature forms, then emphasising how much “doesn’t work yet” becomes a way of protecting against premature extrapolation.

Why it matters now

The substantive meaning of Evans’ remark is that AI should be treated as a major structural shift whose full consequences remain undecided.^1,3,9 The technology is powerful enough that ignoring it is likely to be costly, but immature enough that making large, irreversible bets on specific applications or vendors is risky. Navigating this tension is now a central strategic problem for organisations, investors, policymakers and workers.

For practitioners, the lesson is to separate conviction about direction from confidence about detail. It is reasonable to believe that AI-enhanced systems will permeate most digital workflows, just as networked services and mobile devices did. It is much less reasonable to assume that any particular current configuration – a given model architecture, product category or pricing structure – will survive intact. This distinction enables committed experimentation without dogmatic commitment.

For policymakers, appreciating the “1997” nature of the moment underscores the need for adaptive regulation. Just as early internet rules evolved through trial, error and jurisprudence, AI governance will likely require iterative approaches, focusing first on clear harms while leaving room for positive-sum innovation. Overly rigid frameworks risk locking in today’s imperfect systems; overly lax ones may allow preventable damage.

And for individuals, the comparison serves as a reminder that skill sets, habits and mental models will need periodic revision. Learning to work effectively with imperfect AI tools – much as earlier generations learned to navigate clunky browsers, unstable connections and evolving interfaces – is likely to be more valuable than trying to predict precise future labour-market configurations. The discomfort of using technology that “kind of doesn’t work yet” is not a bug of the current moment; it is a feature of living through the build-out of a new computational substrate.

 

References

1. “The most rational take on AI you’ll hear this year – Lenny’s Podcast – 31 May 2026” – https://www.youtube.com/watch?v=BD3vLtWhT5A

2. AI is like 1997 internet, uncertain value and distribution key – LinkedIn – 2026-06-01 – https://www.linkedin.com/posts/anthony-k-5a883228b_benedict-evans-says-ai-is-exactly-where-the-activity-7467215767323824129-V9AJ

3. Artificial Intelligence Inheriting the Historical Crisis in Psychology – 2022-03-10 – https://pmc.ncbi.nlm.nih.gov/articles/PMC8961441/

4. Benedict Evans: AI Matches Internet Scale, We Are In 1997 Era – Digg – 2026-06-06 – https://digg.com/tech/gadap3fr

5. Artificial Intelligence – an overview | ScienceDirect Topics – 2012-10-30 – https://www.sciencedirect.com/topics/social-sciences/artificial-intelligence

6. The Economics of AI Usage and What’s Next For SaaS – YouTube – 2026-06-08 – https://www.youtube.com/watch?v=ktl8mNiWqMM

7. Analogy and the Roots of Creative Intelligence | The MIT Press Reader – 2025-03-13 – https://thereader.mitpress.mit.edu/analogy-and-the-roots-of-creative-intelligence/

8. A rational conversation on where AI is actually going | Benedict Evans – https://open.spotify.com/episode/5Vqp5z6WshxyfAMBGxHzoh

9. AI Explained with a School Analogy – YouTube – 2025-10-08 – https://www.youtube.com/watch?v=AsipquItRp0

10. A rational conversation on where AI is actually going | Benedict Evans – 2026-05-31 – https://www.lennysnewsletter.com/p/a-rational-conversation-on-where

11. A perfect analogy for AI – YouTube – 2023-08-15 – https://www.youtube.com/shorts/mXa-bS79yq4

12. Why AI Feels Like the Internet in 1997 | Benedict Evans on a16z – 2026-06-08 – https://www.reddit.com/r/PostAI/comments/1u0jbzo/why_ai_feels_like_the_internet_in_1997_benedict/

13. Modality and representation in analogy | AI EDAM | Cambridge Core – 2008-03-14 – https://www.cambridge.org/core/journals/ai-edam/article/modality-and-representation-in-analogy/F567723948DE2E845EDF1E048C87D9EF

14. AI eats the world (Spring 26) [pdf] – Hacker News – 2026-05-19 – https://news.ycombinator.com/item?id=48179021

15. Computational models of analogy – 2011 – WIREs Cognitive Science – 2010-09-20 – https://wires.onlinelibrary.wiley.com/doi/10.1002/wcs.105

16. Presentations – Benedict Evans – https://www.ben-evans.com/presentations

17. Analysis of key AI analogies – EA Forum – 2024-06-29 – https://forum.effectivealtruism.org/posts/3dqjtbtJywQBsC9nA/analysis-of-key-ai-analogies

18. Design, Analogy, and Creativity – IEEE Computer Society – https://www.computer.org/csdl/magazine/ex/1997/03/x3062/13rRUxAASOD