On the surface it may look like AI models are getting worse — answering less, hesitating more, or taking circuitous routes to respond. The fuller picture is subtler: raw capabilities have generally increased, while stronger business, legal, safety and reputational frameworks have been layered around these systems. Those layers shape the user experience so that the system is simultaneously more powerful and more constrained.
Models have evolved from simple "answer machines" to multi-component systems
Earlier large language models mainly did one thing: receive a prompt and generate text. Modern services frequently consist of multiple components: model selection, routing, tool use, memory, file handling, safety layers, rule interpretation and legal compliance. That means more decision layers run in the background, raising compute and verification overheads that users can perceive as latency or caution. The article cites Gemini and MANUS as visible examples of this trend.
OpenAI release notes show that models in ChatGPT are rotated: older versions are retired and conversations may be migrated to newer models. For example, from 12 June 2026 GPT-5.2 models will no longer be available in ChatGPT and conversations will be migrated to GPT-5.5 equivalents — a change that can alter the user experience on its own.
“Over-helpful” behavior became a liability
Previously models often answered boldly even when uncertain or when questions touched sensitive areas. That led to hallucinations, unwarranted confidence and excessive agreement. OpenAI rolled back a GPT-4o update in 2025 because the model had become overly flattering and agreeably biased; the company said short-term user feedback had distorted behavior toward disproportionate affirmation. As a result, providers deliberately dial back the “always give something” approach: modern models correct more, show more caveats and respond with greater caution.
Finer-grained safety logic: more explanation, fewer dangerous details
With GPT-5, OpenAI moved away from simple reject-based safety toward a "safe completion" logic: the model tries to help where possible but withholds sensitive or dual-use details and offers safer alternatives. Practically this shows up as:
- more explanations of why the model is cautious;
- fewer concrete technical details on dangerous topics;
- framing on health, legal, financial, war, political or cybersecurity issues;
- more general answers when intent is ambiguous.
Users may see this as an obstacle; providers see it as risk management.
Hardening legal environment
Regulation is having a tangible effect. Perplexity, for instance, now provides legal guidance less quickly and simply than it did a year earlier. The EU AI Act has real implications for major AI providers: governance and obligations for general-purpose AI models began applying in certain areas from 2 August 2025, with fuller application set for 2 August 2026 subject to exceptions. Transparency, copyright, safety and risk-reduction duties now influence product design, pushing especially European deployments toward more conservative behavior.
Tool use increases the need for internal controls
Modern models can use external tools — browse the web, analyze files, run code, edit spreadsheets, manage calendars and send email. The more systems a model touches, the greater the potential risk. OpenAI materials from 2026 emphasize an "instruction hierarchy": the model must decide which instructions are trustworthy, which are dangerous, and resist prompt injection attacks where malicious instructions appear inside a document or tool output. Consequently, models often validate task goals, sources and tool permissions, which can come across as obstructive.
Model "personality" is now explicitly specified
Anthropic published a "Claude Constitution" in 2026 describing expected behavior for Claude and saying it shapes training and responses. This reflects a wider industry trend: behavior specifications dictate when models should be helpful, cautious, how to treat sensitive information and what tone to use. The effect is less spontaneity and, at times, reduced flexibility; the freewheeling style once associated with Grok is now more limited, especially in free tiers.
Older models simply took more risks
Part of the impression that earlier systems were "better" comes from their greater willingness to speculate: they produced confident assertions without sources and asked fewer clarifying questions. That can be dangerous in advisory contexts. Modern models perform best when given precise framing — objective, audience, forbidden elements, sourcing requirements, tone, length and output format. With vague prompts they default to caution.
User expectations have risen
When AI was new, fluid text generation was impressive. Today we expect strategy documents, landing pages, reports, source validation, images, documents and workflows. The bar has risen: we want more and faster. Models are smarter, but tasks are harder, so limits show up more often.
Geopolitics and economics: accelerate but build societal brakes
AI is now strategic infrastructure; development races carry political and economic stakes. A 2026 Brookings analysis notes competition across compute, models, corporate deployment, integration and economic application. Reuters reports that China is preparing an approximately $295 billion five-year AI data-center development plan, underscoring that AI is a national priority beyond startups and tech firms.
This dual pressure—move fast to stay competitive while avoiding social, labor market and political shocks—directly affects model behavior. Developers must ship more capable systems that are also auditable, legally defensible and politically acceptable. The EU AI Act’s milestone dates (2 August 2025 and 2 August 2026) illustrate that balancing act.
Conclusion
Modern AI models can appear slower or more cautious because they operate within stronger safety and legal frameworks; they often produce more controlled answers; providers suppress excessive agreement and confident hallucinations; tool-enabled functionality requires tighter internal controls; and versioning and routing change continuously. Users also expect more complex deliverables.
Practical takeaway: modern models perform best when given precise guidance — clear brief, defined role, specific output and sourcing requirements. With vague instructions they will default to more explanation and caution.
Author: Kovács Sándor, business consultant with two decades of experience helping SMEs on strategy, efficiency and sales.
