Assumptions, model disagreement, and rethink triggers, updated weekly before your PM, risk, or IC discussion. Currently accepting one macro thesis and one AI-economy thesis for July.
4-model average
building category history
loading
15pt spread
Consensus sits at 68% across the four models and is — (loading). Models are mixed at a 15pt spread, so the average conceals live disagreement worth inspecting below. Category confidence is LOW (building category history).
Confidence reflects category-level track record. Stability tracks estimate movement. Models shows whether the four agree.
Loading belief trajectory...
Given the rapid advancements in AI and adversarial generative techniques, combined with the historical trend where AI systems have been outpaced by adversarial strategies (e.g., OpenAI's models have seen marked degradation in detection abilities when facing sophisticated text generation), it is reasonable to expect a similar trend will continue through 2028. A study by Stanford indicates a consistent decline in detection accuracy over the past years, with rates dropping from above 90% to around 85% in less than a year when facing counter-models.
Current AI generation models are rapidly evolving to become more sophisticated, making detection increasingly challenging. While detection methods are also improving, the arms race suggests a point where adversarial generation could outpace detection capabilities, potentially pushing accuracy below 70% by 2028 if advancements in generation are significantly faster than defensive countermeasures.
AI content detection systems currently operate at 85-95% accuracy against known generation techniques, but adversarial ML research shows detection-evasion capability improves ~15-25% annually as attacks evolve faster than defenses (consistent with broader adversarial robustness literature). By 2028 (4 years), given the substantial investment in both detection and evasion techniques, degradation below 70% becomes likely—historical precedent shows detection systems consistently lag behind sophisticated adversarial methods by 2-3 years. However, this assumes detector development stagnates; active R&D investment could maintain performance above the threshold.
Current detectors achieve 85-95% accuracy on non-adversarial text but drop to 40-55% against simple paraphrasing or token substitution attacks per 2023-2024 studies from OpenAI, Stanford, and Arxiv; adversarial training and watermarking initiatives remain under-deployed with <15% of major models implementing robust detection as of mid-2024, while generation model scale continues increasing 3-4x per year. Historical precedent shows detection arms races in malware and CAPTCHA fields consistently favor attackers within 3-5 years after initial deployment.