The Verification Tax Comes Due

The week’s AI story is not that machines can produce more. They can. It is that every gain in production shifts the bottleneck toward verification: checking tokens, citations, intent, provenance, and the small matter of whether a purported human actually did the work.

At the infrastructure layer, vLLM’s AMD benchmark report describes speculative decoding as a carefully bounded bargain. A draft component proposes several tokens; the target model checks them left to right, retains accepted tokens, rejects the rest, and supplies a replacement when needed. As the post puts it, “Speculative decoding allows vLLM to verify multiple drafted tokens in a single target-model pass.” The target model remains the authority. It is less an autonomous acceleration than an aggressively optimized review queue.

That distinction matters because the benchmark is not selling magic, despite living in an industry that has made a religion of adding the word agentic to a latency chart. On tested AMD configurations, some combinations exceeded 2× baseline output throughput: Gemma 4 MTP reached 2.83× on MATH500 for one Gemma target, DFlash reached 2.79× on HumanEval in another result, and native MTP hit 2.20× on MATH500 for Qwen3.5-122B-A10B. Other settings were below baseline. Model, draft checkpoint, workload and proposal length determine whether the clever shortcut is a shortcut or merely a second model consuming VRAM with tremendous self-esteem.

For CTOs, the operational lesson is wonderfully unglamorous: benchmark the actual request mix, pin the software and driver stack, reserve memory for compatible draft models, and tune num_speculative_tokens rather than treating it as a universal volume knob. The vendor’s own disclaimer is the correct one: “Performance may vary based on configuration, software, vLLM version, and the use of the latest drivers and optimizations.” A faster decoder is a systems result, not a personality trait.

Now move from target-model verification to human verification, where the economics become less charming. In a self-reported experiment, international-law professor Kevin Jon Heller had ChatGPT generate a 6,400-word article with 52 footnotes from a short abstract and instructions intended to reproduce his style. “19 minutes later, ChatGPT gave me the completed 6,400 word, 52 footnote article,” he writes in his account of the exercise.

Heller says five rounds of generation and revision produced a manuscript he thought could be suitable for a good journal, and that he checked its citations without finding invented or miscited sources in this instance. Then came the diagnostic farce. Pangram classified a roughly 500-word passage at no better than 75% AI-written; after Heller manually altered a roughly 700-word section by swapping words, rearranging clauses and changing paragraphs into lists—but without composing sentences from scratch—the detector called it 100% human. “Pangram judged the revised section 100% human-written. Which, really, it wasn’t.”

This is not evidence that every detector is useless, nor that one author’s experiment generalizes across disciplines, models or submissions. It is evidence of a more durable problem: once a system rewards a particular signal, work can be made to resemble that signal. Academic integrity tooling is in danger of becoming an interface-design contest between classifier and copy editor. The detector does not establish authorship; it produces a score whose relationship to authorship can be cosmetically manipulated.

The relevant parallel with speculative decoding is slightly bleak. Both systems rely on a verifier after cheap generation. But the LLM target model can reject a token at machine speed and with a defined distribution. Editors, supervisors and reviewers must assess arguments, sources and intellectual contribution under finite attention. Heller estimates his conventional drafting would have taken about six weeks of a few hours daily; the assisted workflow and enough edits to evade Pangram took about eight hours. That delta is not just a cheating problem. It is a submission-volume problem, a peer-review problem, and eventually an epistemic-surveillance problem.

The evidence that this is not confined to inference clusters and journal workflows arrives in handwriting. Artist Olivia Tai’s project, What We Tell AI, has gathered more than 350 anonymous notes across more than 20 locations since 2025. The confessions cover work, school, dating, breakups, companionship and emotional support. One person reportedly described cloning an ex-partner’s voice for a final conversation; another said they talk to AI more than to everyone else in their life combined.

Tai is careful to call the collection an “ethnographic snapshot,” not a representative survey. Still, broader research cited by CNBC suggests the direction of travel is real: about half of US adults report using AI chatbots, and roughly one in four use them daily. A cited poll found that 27% use chatbots for personal, emotional or social topics; among those users, nearly four in ten said they had told AI things they would not tell another person.

The private chat log is therefore becoming something more consequential than a UI artifact. It is a diary, rehearsal room, advice column, ghostwriter and occasionally the least qualified relationship counsellor in recorded history. Its attraction is obvious: no scheduling, no embarrassment, no reciprocal needs, instant fluency. Its costs are harder to observe precisely because the interaction is private. Tai’s point is that “Almost every person’s history of interaction with an LLM is in private chat logs.”

That privacy also complicates governance. Enterprises can set retention controls and universities can write syllabi, but people increasingly bring the habits formed in intimate consumer chat into work and education. The technical challenge is no longer only model quality. It is the creation of systems people trust with disclosures they may not fully understand as data, and whose persuasive competence can be mistaken for relational competence. A context window, it turns out, can feel remarkably like being heard.

Dordt University’s year-long AI pilot is a comparatively sane institutional response to an unsane baseline: students and faculty were already using these systems independently, so pretending the tools remain outside the gates would be administrative cosplay. Through BoodleBox, Dordt is providing premium access to Claude, ChatGPT, Gemini and Perplexity, while faculty-led monthly discussions take up ethics, education technology and theology. The reported governing policy is admirably direct: “you can use AI to the extent in which your faculty tell you that you can.” Read the report on Dordt’s pilot.

Central access could reduce the absurdity of an institution demanding responsible use while outsourcing practical literacy to whichever free chatbot students happen to find first. It may also let Dordt discuss what users are doing rather than treating AI solely as a disciplinary event. The program aims to make people more critical consumers of the technology; that is better than detector roulette, and much better than issuing stern PDF guidance while everyone quietly pastes assignments into a browser tab.

But buying a multiplexer for several premium models does not solve the hard questions. The article provides no detail on the agreement’s data and privacy protections, cost, participation or eventual results. And a faculty-permission rule is governance only if staff have enough understanding, time and consistency to make permission meaningful. Otherwise it becomes a distributed policy engine with incompatible versions, missing documentation and occasional theological merge conflicts.

The coherent strategy across this issue is not abstinence, worship or a detector bolted to the LMS like a smoke alarm in a server rack. It is to make generation cheap where it should be cheap, and preserve accountable review where it cannot be automated away. Speculative decoding succeeds because the target still verifies the draft. Institutions should take the hint. The question is not whether the machine drafted it. The question is who is willing—and competent—to own what survived review.