When Does AI Reduce Perceived Workload? A Moderator-Focused Review of NASA-TLX Evidence in AI/LLM Interfaces

Research basis: Two Perplexity research passes (2026-08-03/04), independently spot-checked against primary sources across three verification rounds (14 sources fetched directly, not taken on the synthesis's word). Two sources were found cited backwards during verification; those corrections are folded in below, not hidden.

---

Why this article exists

The evidence that AI assistance reduces NASA-TLX-measured workload is not in serious dispute — we have it from independent studies spanning chatbots, copilots, tutoring, swarm control, and clinical communication. What's not well established, and what most secondary coverage skips past, is when it doesn't, or when a lower TLX score is hiding a cost that TLX itself can't see. That's the question this piece answers, moderator by moderator, with each claim graded by how directly it was tested rather than presented as uniformly “proven.”

A note on method, since it matters for how much weight to put on what follows: every quantitative claim below was checked against its primary source directly — not accepted from the research synthesis that surfaced it. Two claims turned out to be backwards relative to their own source paper (flagged explicitly where they occur). Two more sources turned out to sit in low-credibility venues despite numerically accurate reporting. One pillar claim came from a 10-person pilot study its own authors call preliminary. None of that invalidates the overall picture, but it changes which specific claims can carry weight in a client deliverable and which can't.

It's also worth stating up front that NASA-TLX itself is not beyond question as an instrument. Babaei, Dingler, Tag & Velloso's 2025 review in the International Journal of Human-Computer Studies found “a lack of convergent validity and sensitivity of MWL subjective scales in HCI tasks” and recommends caution when using NASA-TLX as a default evaluation tool. Kosch et al.'s 2023 ACM Computing Surveys survey goes further, describing NASA-TLX's dominance in HCI research as closer to “academic tradition” than a reasoned methodological choice — “the primary reason behind the use of NASA-TLX appears to be community convention” rather than demonstrated superiority over alternatives. Everything below should be read against that backdrop: we're asking when a contested instrument moves, not treating its movements as ground truth.

---

1. Task structure is the best-evidenced moderator

The clearest primary evidence for a moderator — tested as an actual statistical interaction, not inferred from separate averages — comes from a Microsoft study of Copilot in Word across four task types (Russell, Shah, Blaney, Amores, Czerwinski & Jacob, Neural and Cognitive Impacts of AI, arXiv:2506.04167):

“The Copilot condition resulted in overall lower TLX scores (F₁,₁₃₃=60.42, p<0.001, ϵp²=.31)... CONDITION×TASK demonstrated a strong effect (F₃,₁₃₃=7.07, p<0.001, ϵp²=.12).”

That interaction term is the key fact: task type doesn't just correlate with workload, it changes the size of Copilot's effect. Structured, objective tasks (SAT-style reading comprehension, event planning) showed significant TLX reduction plus better performance. A constrained-creative task (poem writing) showed TLX reduction and higher enjoyment (1.80 vs 0.62, p=0.005) with no change in output quality. A highly subjective, episodic self-reflection task showed no TLX benefit at all (t₁₃₃=0.17, p=0.864), alongside distinct prefrontal activation on fNIRS relative to the other tasks.

Grade: established. This is a formally tested moderator with a significant interaction term, not a subgroup comparison dressed up as one.

A second data point, weaker in design but consistent in direction, comes from a comparison of chatbot-assisted vs. web-browsing research for essay writing (Marois et al., Chatbot Memory, HFES Proceedings, N=59): mental demand was significantly lower with the chatbot (p=.046), but no other TLX subscale and no memory/essay performance measure differed. Task structure here is held constant (a self-regulated learning task) — the finding says AI can lower one specific dimension of workload without moving the others, which is itself evidence that “workload” isn't a single thing AI uniformly discounts.

2. Expertise as a moderator — thin evidence, and it doesn't point where intuition suggests

The working hypothesis going in was: novices benefit more from AI (it substitutes for missing knowledge), experts see added workload from verifying AI output. The only direct comparison we found — a between-subjects study of generative AI (Galileo) vs. Figma for UI/UX design tasks, novices and experts analyzed separately (Shahzad, Daud & Mughal, IJIST 8(2), 2026) — points the other way:

Novices: TLX 2.22±0.53 (AI) vs 3.25±0.70 (Figma). Experts: TLX 2.42±0.64 (AI) vs 3.79±0.52 (Figma).

Both groups show lower TLX with AI, and the absolute reduction is larger for experts (≈1.37 points) than novices (≈1.03 points) — the opposite of the verification-burden hypothesis. But: no formal AI×expertise interaction was reported (this is a subgroup comparison, not a moderator test), and — this matters — the venue is weakly vetted. IJIST/50Sea shows a ~6-week submission-to-publication turnaround and indexing only by low-bar aggregators (no genuine Scopus or DOAJ listing, despite an earlier automated summary claiming Scopus indexing, which we could not substantiate). The same study also found creativity scores were higher with traditional tools for both groups (novices 3.94 vs 3.27; experts 4.17 vs 3.37) — so even where TLX drops, output quality by this measure doesn't follow it down.

Grade: suggestive, not established. The numbers are internally consistent and the underlying data checks out against the paper — but the paper itself is not a source we'd want to hang a client recommendation on alone, and the finding contradicts the intuitive hypothesis rather than confirming it. Flag this as an open question, not an answer.

A frequently-cited third data point (a JETIR paper claiming digital-literacy correlates with lower TLX, r≈−0.25) failed verification twice: the correlation is stated three times in the paper's text but its own scatter-plot figure prints r=−0.28 for the same result, and a related claim about frustration rising with response ambiguity turned out not to be a reported finding at all but an inference stitched from two disconnected numbers. We're dropping this source from the evidentiary base entirely rather than citing it with caveats — it's failed independent verification on two separate claims now.

3. Trust and verification burden — the evidence contradicts the intuitive hypothesis

This is the section where the corrections matter most.

The working hypothesis: adding human oversight/verification to an AI system should show up as added workload on specific TLX subscales (frustration, mental demand), even if overall workload drops. One frequently-cited source seemed to confirm this — a four-condition UAV swarm-teaming study (Ji, Hu, Zhang & Chen, LLM-CRF, arXiv:2511.04042) comparing manual control, direct LLM control, LLM-CRF without human feedback, and the full LLM-CRF with human-in-the-loop verification. The original research synthesis reported the full (human-verified) condition as higher workload than the no-feedback autonomous condition — i.e., oversight costs something.

That's backwards. The paper's own results table says the opposite:

Config	Mission Success	NASA-TLX
B1 Manual	87.0%	71.2±9.3
B2 LLM-Direct	11.0%	68.5±13.7
B3 Ours, no feedback	62.0%	42.8±8.1
Ours, Full (human-in-the-loop)	94.0%	28.3±6.2

The full framework, with human verification, has the lowest workload and the highest mission success of all four configurations — lower than the no-feedback version, not higher. This is confirmed twice within the paper itself, independently of the table: the Conclusion states plainly that mission performance “was also maintained with the operator's cognitive load (NASA-TLX) of 28.3%, confirming the framework's success in alleviating the mental burden of complex swarm management” — assigning 28.3 to the full, human-verified framework, matching Table 3.

One paragraph elsewhere in the paper's Results Analysis section does state the reverse (“the NASA-TLX score which increased from 28.3 to 42.8”) when describing the same B3-to-Full comparison — a genuine internal inconsistency in the source, not a misreading on our part. We flag this explicitly because a third-party summary of this same paper, checked during this article's review, reproduced that exact reversed reading as fact — meaning anyone verifying this claim against a secondary review rather than the primary table and Conclusion is likely to hit the same error we did. Table 3 and the Conclusion agree with each other and are the paper's two most explicit, considered statements of the result; the single contradicting sentence sits alone against both.

What actually holds up on this moderator: two studies show AI can lower workload on some subscales while a different signal moves in the opposite direction, but not in the “verification adds mental demand” pattern hypothesized. A VR+ChatGPT nursing OSCE trial (Dai, Gong & Ma, randomized crossover, N=65) found overall TLX dropped from 52.9±13.1 to 47.2±12.5 (p=0.001), driven mainly by frustration (−8.1, p<0.001) and temporal demand (−6.4, p=0.004) — not mental demand. And a human-robot collaboration study with LLM integration (Su, Cheng, Lu, Qing, Jung & Xu, Applied Ergonomics 134:104736, 2026) found lower subjective TLX (mental demand, effort, frustration) and higher trust and elevated GSR arousal simultaneously:

“GSR data indicated elevated physiological arousal, possibly suggesting increased engagement or positive emotional activation.”

Note the authors' own framing: they read the elevated arousal as positive engagement, not masked stress. The earlier research synthesis presented this as an unresolved subjective/objective tension; the paper itself doesn't support that framing, and neither should we.

Grade: the intuitive hypothesis (oversight adds workload) is not supported by the evidence we have — if anything, the one study that directly tests it shows the opposite. This is worth stating plainly in any client-facing use of this material: don't assume adding a human-verification step to an AI workflow will show up as added TLX burden. The available evidence says it might not.

4. Interaction modality — chat/conversational interfaces reduce mental demand specifically, not workload generally

Three comparisons here, of varying strength. The Chatbot Memory study above shows mental demand down with nothing else moving. A clinical communication study (Fudan University urology department, preoperative patient communication) shows physician NASA-TLX dropping from a median of 54.0 [IQR 35.0–73.0] to 37.0 [IQR 27.0–46.0] with LLM-assisted communication, alongside shorter communication time (19.4 to 10.8 minutes). And an earlier, smaller enterprise study (Schmidhuber, Schlögl & Ploder, arXiv:2111.01400) found software-only users perceived 32% higher cognitive load than chatbot-assisted users (28.6 vs. 21.6), alongside 60% higher productivity and higher task accuracy (86.2% vs. 63.8%) for the chatbot group, despite averaging about four minutes longer per task — chat modality helping accuracy and workload simultaneously even at the cost of some speed. That study is n=22 in a single Austrian B2B setting, so it's a data point, not a generalizable claim on its own.

All three point the same direction: conversational/AI-mediated interaction reduces workload relative to manual search or manual dialogue, but the mechanism looks like reduced search/navigation and time pressure specifically, not a uniform discount across all workload dimensions.

Grade: established for the “reduces specific subscales, not overall workload uniformly” pattern; not established as a formally-tested moderator (these are condition comparisons, not interaction tests).

5. Stakes/consequence level — real variation, but not a tested moderator

Effect sizes vary widely across contexts: the corrected LLM-CRF swarm study (full framework, high-stakes SAR) shows a 42.9% relative TLX reduction with substantially improved mission success; the Fudan clinical study shows a ~31% reduction; the nursing OSCE shows an ~11% reduction; a dental-education study (Bhadila, Bahdila, Saber & Alyafi, Frontiers in Education, N=132) shows TLX dropping from a median of 41.7 [IQR 30.0–62.5] to 21.7 [IQR 8.3–46.7] (p<0.0001) alongside improved performance (median score 13 vs 11, p<0.0001, adjusted OR=1.67 for correct answers).

No source in this set formally tests “stakes” as a moderator with an interaction term — this is cross-study comparison, and should be presented as a hypothesis worth designing a real test for, not a finding.

6. Where NASA-TLX and objective measures diverge

This is the section that most directly bears on whether we should trust NASA-TLX as a standalone signal, and it's mixed:

• Convergence: a driver-dialogue study (Fredriksson, Yaici, Lam, Konigsmann & Edlund, IWSDS 2026) found NASA-TLX mental demand rising 136% from baseline to high-complexity conditions, alongside a 60% slowdown in Detection Response Task reaction time and a hit-rate drop from 93.1% to 59.8%, with TLX correlating strongly with both DRT reaction time (r=0.92) and collision frequency (r=0.85). This is the strongest convergence evidence in the set — but it's an n=10 pilot study that its own authors explicitly label preliminary, not a generalizable finding. It shouldn't be presented as a settled pillar without that caveat attached every time it's cited.
• Non-convergence: the Copilot study found no significant fNIRS or HR/HRV/EDA changes attributable to Copilot despite the TLX and performance differences reported above — physiological measures were largely insensitive to the intervention that moved subjective workload.
• Partial divergence, reframed: the human-robot GSR study (elevated arousal alongside lower TLX) looks like divergence at first glance but the authors interpret it as engagement, not conflicting signals — see Section 3.
• Outcome divergence: the Galileo/Figma study shows lower TLX and faster completion coexisting with lower creativity ratings — a case where reduced perceived workload doesn't track with a quality outcome, though this sits in the same weakly-vetted venue flagged in Section 2.

Grade: genuinely mixed, and that's the honest finding. NASA-TLX tracks objective measures well in at least one (small, preliminary) study and not at all in another (well-powered, physiological) one. Anyone using NASA-TLX as a sole outcome measure for AI-interface evaluation should pair it with something else — performance data at minimum — precisely because we can't yet predict which case we're in.

---

Synthesis: what's actually load-bearing

The most useful takeaway for our own UX-research design work: if we ever run NASA-TLX on an AI-assisted workflow, (1) expect the effect to depend heavily on task structure, so segment by task type rather than reporting one aggregate score; (2) don't assume a human-verification step will show up as added burden — test it, don't assume it; and (3) pair TLX with at least one objective measure (performance, error rate, or physiological if feasible), because the one study that tested convergence directly was underpowered and the one that was well-powered found no convergence at all.

---

Sources