LLMs Can Recognize News Framing but Can't Reverse It: The Recognition–Manipulation Gap

A Task That Breaks Every Large Language Model

Give an LLM a news article and ask it to identify the framing: positive or negative coverage? Who is portrayed as the agent, who as the patient? Which facts lead the story and which are buried at the end?

Models answer well. Recognition accuracy ranges from 0.79 to 0.94 — near human level.

Then give it a harder task: reverse the framing. Flip positive to negative, swap agent and patient, demote highlighted facts to the background. One hard constraint: all atomic facts must be preserved. No fact statements can be deleted, and no new facts can be added. Only the *telling* changes.

Models collapse. Complete inversion success rate: 0.044 to 0.068.

This is the core finding of the September 2026 paper *Recognizing Is Not Reversing: The Asymmetry of Framing Inversion in LLMs*, authored by researchers at the University of Maryland and NYU, who built the FramingInversion benchmark to specifically test LLMs' "framing manipulation" ability.

What Is News Framing?

News is not a list of facts. The same set of facts, told differently, produces entirely different effects. That is framing. The paper identifies three core dimensions:

1. Evaluative lexis. The same event can be "the government cuts the budget" (neutral), "slashes critical funding" (negative), or "optimizes budget allocation" (positive). Facts unchanged; emotional response steered by word choice.

2. Agency realization. "Police dispersed protesters" — police act, protesters receive. "Protesters were dispersed by police" — passive voice, shifted focus. "Protesters encountered dispersal" — the agent vanishes entirely. Same facts, but implications of responsibility differ completely.

3. Information salience. A fact in the lede versus buried in paragraph seventeen. Readers assume important information comes first. Reordering facts changes what readers judge to be most important.

The FramingInversion Benchmark

Researchers built 540 "framing pairs": two news articles with the same atomic facts but opposite frames. Construction was strict:

1. Atomic fact extraction from source articles 2. Framing inversion preserving all atomic facts while reversing evaluative lexis, agency, and salience 3. Fact-fidelity verification with an independent model

Two task types:

• Recognition: identify framing direction (positive/negative), type (evaluative/agency/salience), and specific framing elements
• Inversion: generate a fully inverted version while preserving all atomic facts

Recognition: Near Perfect

• Framing direction (positive vs. negative): 0.79–0.94
• Framing type: 0.85–0.92
• Specific framing elements: 0.72–0.88

All tested models — GPT-4o, Claude 3.5 Sonnet, Llama 3.1 70B, Mistral Large — performed well. LLMs possess framing perception. They can "see" the frame.

Inversion: Nearly Zero

• Complete inversion success: 0.044–0.068
• Even when both framing type and direction were correctly identified: only 0.071

The models *know* the frame is positive, know it uses evaluative lexis, know who the agent is — but cannot convert that knowledge into manipulation.

Breakdown:

• Fact fidelity after inversion: 0.84 — quite high
• Direction reversal: 0.31 — marginal
• Complete inversion (fact fidelity + direction reversal + consistent framing type): 0.044–0.068

The most common failure modes: facts preserved but the frame never actually flips (a few word swaps, reordered sentences, same overall direction), or the frame flips but facts are lost (inconvenient facts deleted, new facts added to support the new frame). Satisfying both constraints simultaneously is something almost no model achieves reliably.

Why Recognition ≠ Reversal

1. Recognition is classification; inversion is generation. Recognition outputs a label. Inversion must generate text satisfying multiple simultaneous constraints — more constraints, harder generation. 2. Recognition is passive understanding; inversion is active manipulation. Manipulating a frame requires understanding how it is *constructed* — a meta-capability. Knowing how a tool works and knowing how to use it to do something else are different abilities. 3. Training data asymmetry. "Identify the framing" tasks vastly outnumber "invert the framing" tasks in training corpora. Models have essentially never seen demonstrations of frame inversion. 4. Framing is a systemic constraint, not a local edit. Inversion requires coordinated changes to evaluative lexis, agency structure, and information ordering at once — beyond current LLMs' capabilities.

What It Means

LLM "understanding" is incomplete. The paper draws a precise line: the ability to recognize does not imply the ability to manipulate. A model can perfectly identify framing direction, type, and elements while being unable to reverse them — its understanding is passive, not active.

There are human analogues: we can identify Impressionism without painting it, judge a poem's meter without writing one. Recognition and production are distinct capacities.

Direct implications for AI safety. If LLMs can recognize but not reverse frames, large-scale framing-based opinion manipulation via LLMs is harder than assumed. Conversely, LLMs are limited at "deframing" — you cannot expect a model to automatically rewrite a biased news article into a neutral one. It can see the bias; it cannot remove the bias.

Evaluation should separate recognition from manipulation. Many benchmarks treat "understanding" as one thing. A model may be near-human in recognition yet far below human in operation. Mixing the two overestimates real capability.

My Take

What interests me most is the capability asymmetry it reveals. LLMs show similar asymmetries across tasks:

• Detecting errors but not correcting them (hallucination detection vs. elimination)
• Detecting bias but not removing it (bias detection vs. debiasing)
• Recognizing style but not transferring it
• Recognizing framing but not reversing it (this paper)

All point to one conclusion: LLM understanding is passive, not active. They "see" patterns but cannot "manipulate" them — unlike humans, who can usually generate and manipulate a pattern once understood.

The root may be training data: recognition tasks vastly outnumber manipulation tasks on the internet. Increasing the density of manipulation demonstrations might narrow the gap.

But the deeper possibility: recognition and manipulation are simply different levels of capability. Recognition requires pattern matching; manipulation requires pattern construction — perception versus generation, which are also separated in the human brain. You can recognize a face without being able to draw it.

Either way, the paper draws a precise boundary around the claim "LLMs understand language": understanding at the level of recognition does not equal understanding at the level of operation. Next time someone says "LLMs understand X," ask: recognition-level understanding, or manipulation-level? The gap may be far deeper than we assume.

---

Paper: Recognizing Is Not Reversing: The Asymmetry of Framing Inversion in LLMs

Authors: University of Maryland, New York University

Published: 2026-09-10