Why AI Detection Is Harder Than 'Real or Fake'

As generative AI floods the internet with synthetic text, images, and video, the demand for reliable detection tools has never been higher. But according to Max Spero, CEO of AI detection startup Pangram, the popular framing of detection as a simple binary question — "real or fake?" — badly underestimates the technical complexity involved. In a recent TechCrunch interview, Spero unpacked why building trustworthy detectors is one of the thornier problems in the synthetic media landscape.

Beyond Binary Classification

The instinct to treat AI detection as a yes-or-no output is understandable but misleading. Spero argues that reducing detection to a single boolean value obscures the probabilistic reality of how these systems actually work. Modern detectors don't "know" whether a piece of content is AI-generated — they estimate a likelihood based on statistical patterns, token distributions, and stylistic fingerprints that generative models tend to leave behind.

That distinction matters enormously in practice. A detector that confidently labels a human-written essay as "AI-generated" can have serious real-world consequences — a student falsely accused of cheating, a journalist wrongly flagged, or a job applicant unfairly screened out. The cost of a false positive is often far higher than the cost of a missed detection, which forces detection companies to think carefully about calibration and thresholds rather than chasing raw accuracy numbers.

The False Positive Problem

One of the central challenges Spero highlights is minimizing false positives while maintaining useful detection rates. This is a delicate balancing act. Push a detector toward high sensitivity, and it will start flagging legitimate human work. Tune it toward caution, and sophisticated AI-generated content slips through undetected.

Pangram's approach emphasizes reducing false positives to build trust with users who rely on the tool for high-stakes decisions. This is a meaningful technical stance: many first-generation AI detectors earned reputations for unreliability precisely because they produced confident-sounding verdicts on ambiguous inputs. When a detector cries wolf too often, institutions stop trusting it — undermining the entire value proposition.

An Arms Race Against Model Drift

Another dimension Spero emphasizes is that detection is a moving target. Every new frontier model — from updated versions of GPT to Claude, Gemini, and open-weight alternatives — shifts the statistical signatures that detectors rely on. A model that generates more "human-like" text, or one that a user has explicitly prompted to vary its style, can evade detectors trained on older output distributions.

This creates a continuous arms race. Detection systems must be regularly retrained and updated to keep pace with the models they aim to identify. Unlike a static classification problem, AI detection requires ongoing data collection, adversarial testing, and model maintenance. Detectors that were state-of-the-art a year ago may be substantially degraded against today's generation systems.

Why This Matters for Digital Authenticity

Spero's insights extend well beyond text detection into the broader synthetic media ecosystem. The same fundamental challenges — probabilistic outputs, false positive costs, and constant model drift — apply to deepfake video detection, voice clone identification, and AI image forensics. As generative tools improve, the gap between authentic and synthetic content narrows, making detection harder across every modality.

The lesson for anyone building or deploying authenticity tools is that reliability and calibration matter more than headline accuracy figures. A detector that produces a defensible confidence score, communicates uncertainty honestly, and minimizes harmful false accusations is far more valuable than one that offers a crisp but unreliable verdict.

As AI-generated content becomes ubiquitous, the industry may need to shift from asking "is this real or fake?" toward more nuanced frameworks — provenance tracking, content credentials, watermarking, and probabilistic assessments used in combination. Detection alone, Spero suggests, is unlikely to be a silver bullet. Instead, it will be one layer in a broader stack of tools that together help preserve trust in digital media.

For an ecosystem grappling with the erosion of the line between human and machine-generated content, Spero's candor about the limits and complexities of detection is a useful reality check. The challenge isn't just building better classifiers — it's building systems people can actually trust.

Stay informed on AI video and digital authenticity. Follow Skrew AI News.