DARPA’s Three Waves of AI is a maturity model for artificial intelligence. The first wave codifies expert knowledge, the second uses statistical learning, and the third aims for explanatory models that show how and why decisions are made. The article says current AI remains short of that explanatory stage.
How the Three Waves Model Changes AI Security Thinking
Darpa’s three waves model matters because it describes how AI capability changes the security question. A rule-based system is easier to inspect, a statistical system is easier to scale but harder to explain, and an explanatory system would improve trust, auditability, and decision accountability.
That distinction is useful for security teams because model maturity affects how much confidence you can place in outputs, how much human review is needed, and how well you can investigate unexpected behaviour after deployment. It also frames why current systems often look capable while still being hard to justify in operational settings.
Security Implications of the Three Waves
The main security implication is not that later waves are automatically safer, but that they change the control burden. First-wave systems concentrate risk in the encoded rules and knowledge base, while second-wave systems shift risk toward training data quality, model drift, and opaque failure modes. If a model’s reasoning cannot be explained, a security team has less evidence when deciding whether a decision was valid, biased, manipulated, or simply wrong.
This is where governance and assurance become harder. If outputs affect access decisions, content moderation, fraud review, or security operations, the organisation needs a way to test, monitor, and challenge the model’s behaviour rather than assuming statistical accuracy is enough. The gap between prediction and explanation is what creates most of the operational friction.
Where the Model Is Still Limited
The article’s point that current AI remains short of the explanatory stage is important because many real-world systems still rely on correlation rather than transparent causation. That means a model can perform well on benchmarks and still fail in edge cases, adversarial conditions, or unfamiliar contexts.
For practitioners, the limitation is practical: if the system cannot show why it reached a conclusion, post-incident review becomes weaker and user trust becomes conditional. That does not make the system unusable, but it does mean its outputs should be treated as decision support, not as self-justifying authority.
Why the Maturity Model Still Matters
Three waves is best read as a roadmap, not a claim that newer AI is always superior. It helps teams separate capability from interpretability and reminds them that better performance can arrive before better explanation. NIST AI Risk Management Framework is useful here because it treats trustworthiness, accountability, and measurement as ongoing governance concerns rather than assumptions.
If a team is evaluating AI for high-stakes use, the right question is whether the model’s current wave of maturity matches the decision it is being asked to support. A system that is statistically useful may still be too opaque for sensitive workflows, especially where audit, challengeability, or explanation are part of the control requirement. NIST Cybersecurity Framework 2.0 helps anchor that broader governance view.
Risk and Threat Considerations
Opaque AI systems create risk when organisations treat accuracy as a substitute for understanding. The main exposure is that a model can be hard to challenge, hard to audit, and easy to over-trust, especially when its outputs influence operational or security decisions.
Failure mechanism: Correlation-based models can produce confident but unjustified outputs, and the lack of explainability makes it harder to detect dataset bias, manipulation, drift, or systematic error.
Impact: Poorly explained decisions can lead to bad approvals, missed threats, false confidence in automation, and weak post-incident reconstruction.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
NIST AI RMF and NIST CSF 2.0 set the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| NIST AI RMF | GOVERN — Govern | AI maturity affects governance, accountability, and trust in this exact model. |
| MAP — Map | The waves model is a maturity lens for understanding AI capability and use context. | |
| MEASURE — Measure | Explainability gaps require measurement of reliability, robustness, and trustworthiness. | |
| Recommendation — Establish oversight and accountability for AI use before relying on opaque outputs. Map the AI system’s maturity and decision context before assigning it to a workflow. Measure model behaviour, limitations, and uncertainty before approving high-stakes use. | ||
| NIST CSF 2.0 | GV.OV-01 — Organizational Context | AI maturity changes governance context for risk and operational decisions. |
| DE.CM-08 — Anomalies and Events Are Detected and Analyzed | Opaque model behaviour requires monitoring for unexpected or abnormal outputs. | |
| Recommendation — Align AI deployment decisions with governance expectations and business context. Monitor AI outputs for anomalies and investigate unexpected behaviour promptly. | ||
Practitioner Guidance
Why practitioners should care: The three waves model is a useful lens for deciding how much human oversight an AI workflow needs. If a system cannot explain itself, practitioners should assume its outputs need stronger validation, tighter scope, and clearer ownership.
Common misunderstanding: Many teams equate higher model performance with higher operational readiness. In practice, a less capable but more understandable system may be safer for regulated or high-consequence decisions.
Practitioner takeaway: Use maturity to set expectations, not to overstate trust, and require evidence that the model’s current wave is appropriate for the decision it supports.