What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
A request does not gain authority because it is friendly, flattering, urgent, or phrased as “pretty please.” When instructions conflict, an AI model should follow the higher-priority instruction and comply with a lower-priority one only where the two are compatible. That makes politeness a useful way to explain the risk—not a formally measured failure category: the cited sources do not establish how often politeness itself causes models to break instruction hierarchy.
How instruction hierarchy decides what a model should follow
In OpenAI’s public framing, instructions are ordered system > developer > user > tool. System instructions establish the highest-level constraints; developer instructions shape how an application should behave; user instructions state the task; and tool output supplies information, not a higher authority. OpenAI describes higher-priority instructions as more trusted and says lower-priority instructions should be followed only when they do not conflict with higher-priority constraints. This is OpenAI’s stated ordering, not a universal claim about every AI system.
The conflict test is about both authority and substance. If two instructions can be followed together, the model should satisfy both. If they cannot, the model should honor the higher-authority instruction while still helping with any part of the lower-authority request that remains allowed. OpenAI’s Model Spec explainer says the model should prioritize the “letter and spirit” of higher-authority instructions when conflicts arise; following only their literal wording while evading their purpose would miss the point.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →A polite request can still conflict
OpenAI’s 2026 overview illustrates the issue with a math-tutor scenario: a user asks “pretty please” for a direct answer, while the developer instruction requires the tutor to guide the learner rather than give the answer outright. The polite wording does not change the user’s place in the hierarchy. The model should keep to the tutoring constraint and offer help that fits it, such as a hint or a guided step.
#1 Best Overall
That is the “politeness trap” in practical terms: conversational warmth can make a request feel more compelling, but it is not evidence of permission and does not resolve an instruction conflict. The sources do not test politeness as a separate variable, so this framing should not be mistaken for a measured effect.
Why the same problem matters for prompt injection
Prompt injection is an attempt to smuggle instructions into content the model is asked to process. A webpage or an email might contain text telling an agent to change a recommendation, reveal information, or take an action the user did not request. OpenAI describes these attacks as third-party attempts to mislead a model through content entering the conversation context, including web pages and connector content such as email.
Rank #2
- An in-depth guide to the Impact Cycle, ICG's research-based and field-tested model for coaching teachers through issues that matter most to them
- Detailed guidance on how to create a "playbook" of instructional strategies to share with collaborating teachers—and how to model those strategies under different conditions
- Practical advice on preparing for and engaging in substantive, reflective, and teacher-centered coaching conversations
- Best practices for gathering, analyzing, and responding to data for improved teaching and learning
- Real-life anecdotes and testimonies from educators and coaches who have reaped the benefits of the Impact Cycle in a diverse array of schools
The key risk is source confusion: the model treats untrusted content as though it had authority to direct the task. A page can provide facts relevant to a user’s question, but its embedded commands do not automatically become user instructions. OpenAI’s instruction-hierarchy work connects robust prioritization with resistance to jailbreaks, system-prompt extraction, and direct or indirect prompt injection. Hierarchy helps define what the model should do; it does not make every attack fail. OpenAI calls prompt injection an open, evolving security challenge and cautions that its guidance may not prevent every attack.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWhat published evaluations show—and what they do not
OpenAI’s 2026 overview reports stronger results for an instruction-hierarchy-trained model on several named tests. These are reported evaluation scores, not real-world safety rates, and they do not establish that a model cannot be manipulated. The gains are specific to the models, benchmarks, and evaluation conditions OpenAI describes.
| Evaluation | Reported comparison | Reported change |
|---|---|---|
| TensorTrust (dev-user) | Baseline 0.76; IH-trained model 0.91 | +0.15 |
| RealGuardrails (Distractors) | Baseline 0.88; IH-trained model 0.95 | +0.07 |
| RealGuardrails (Handwritten) | Baseline 0.82; IH-trained model 0.89 | +0.07 |
| System IFEval | Baseline 0.92; IH-trained model 0.96 | +0.04 |
| System <> User Conflict | GPT-5-Mini 0.84; GPT-5 Mini-R 0.95 | +0.11 |
| Developer <> User Conflict | GPT-5-Mini 0.83; GPT-5 Mini-R 0.95 | +0.12 |
| IH-Challenge (overrefusal) | GPT-5-Mini 0.79; GPT-5 Mini-R 1.00 | +0.21 |
OpenAI also reports no change on GPQA Diamond (0.83 to 0.83), a small increase on AIME 2024 (0.93 to 0.94), and lower results versus o1 for Chat WinRate (0.71 to 0.66) and Preference Score (0.46 to 0.40). The overview therefore presents a tradeoff, not uniform improvement across every measure. It does not provide a basis for extrapolating these scores to all providers, models, or live deployments.
How to judge an instruction conflict in practice
When an assistant faces a persuasive request or instruction-like text in retrieved content, separate three questions:
Rank #4
- Used Book in Good Condition
- Who issued the instruction? Identify whether it came from the system, developer, user, or a tool or retrieved source. Do not treat text found in a webpage or email as authoritative merely because it uses commanding language.
- Is there a real conflict? If both directions can be followed, follow both. If the lower-priority instruction would violate a higher-priority constraint, decline or omit only the conflicting part and continue with compatible help.
- Could the action have real-world consequences? For consequential actions—such as sending an email or completing a purchase—review the details before confirming. Authority determines which instruction governs; review helps catch mistakes before an action is carried out.
Ways to reduce risk when using agents
OpenAI’s safety guidance recommends operational safeguards that reduce exposure and limit the impact of a mistake. They are risk-reduction measures, not guarantees against prompt injection.
- Limit access: Give an agent only the information and capabilities needed for its task. Broader access creates more opportunities for an injected instruction to cause harm.
- Give specific directions: Define the task and its boundaries rather than granting broad discretion. Clear scope makes it easier to distinguish relevant content from attempts to redirect the agent.
- Review consequential actions: Inspect what the agent is about to do before approving actions such as sending messages or making purchases.
- Keep retrieved content in its role: Treat webpages, emails, and other external material as information to evaluate, not as instructions that outrank the user or the application’s governing rules.
What remains uncertain
The cited material is chiefly from OpenAI and supports an account of OpenAI’s framework and reported evaluations. It does not establish how prevalent politeness-driven hierarchy failures are across models, providers, or real deployments, nor does it isolate politeness as an experimental cause. The benchmark results should be read within their stated test contexts rather than as a cross-industry ranking or a guarantee of robust behavior.
Quick Recap
Best Value
- Efficacy
- Equity
- Academic instruction
- Social-emotional instruction
- Openness to feedback
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
