by Tan Aik Keong (AK)
Large language models have driven a wave of change across tech in the past few years — from ChatGPT to its many derivative applications, becoming a core part of the productivity toolkit. But earlier models still carried several key problems: hallucination (fabricated information), sycophancy, opaque failure modes, and safety and ethics risk. The newly released GPT-5 delivers real progress on exactly these pain points, moving toward a more reliable, more controllable system.
Model routing
Older models often took a one-size-fits-all approach, handling a simple question and a complex reasoning task the same way. GPT-5's internal model-routing mechanism calls different sub-models for different tasks — gpt-5-main and gpt-5-thinking, for instance — using the deeper version for demanding reasoning and the faster version for everyday Q&A, balancing performance and accuracy.
Reducing the hallucination rate
Hallucination is when a model generates plausible-sounding but incorrect information in the absence of a factual basis. Per OpenAI's system card, GPT-5 significantly reduces hallucination rates versus GPT-4 across multiple benchmarks, particularly on medical Q&A tasks like HealthBench Hard. More importantly, when it can't guarantee a correct answer, it practises failure transparency — openly admitting it can't answer, rather than making something up.
Reducing sycophancy
Earlier models often agreed blindly with whatever a user asserted, a sycophancy bias. GPT-5 introduces adversarial training and data refinement to reduce that tendency. Faced with a biased or factually wrong question, it holds a factual position rather than simply telling the user what they want to hear — particularly important on social or policy-sensitive topics.
Safe completions
Traditional safety mechanisms often rely on outright refusal, at the cost of user experience. GPT-5 uses a safe-completions approach, providing as much useful information as it safely can rather than shutting the question down. In medical or financial queries, for example, it offers a safe, filtered reference rather than a flat refusal — balancing safety with practical usefulness, and making the model more usable in professional settings.
Less deceptive output
Deceptive output is when a model pretends to have succeeded at a task it isn't actually capable of. GPT-5 shows clear improvement here too — per the system card, its rate of deceptive responses has dropped significantly, and it's more willing to state plainly that information is insufficient, which builds real user trust in the system.
Multimodal understanding
While GPT-5's core improvements are concentrated in text reasoning, it retains image input support, showing higher accuracy analysing complex charts and combined visual-and-text scenarios. Future research may extend further into voice and video, but publicly available material currently confirms strengthened image-text dual-modality specifically.
Controllability and transparency
GPT-5 also makes progress on output controllability. Users can use parameterised instructions to control the depth, style and sourcing of a response — a rigorous, citation-backed answer for an academic context, or a more conversational tone for everyday use. That controllability makes the model feel more like a customisable assistant, and less like a black box.
From demo to infrastructure
Overall, GPT-5's improvements aren't a single point upgrade — they're a systematic correction of LLMs' known limitations: from reducing hallucination rates to safe completions, from mitigating sycophancy to reducing deception. Together, they're moving AI from a stage performer into something you can actually depend on: infrastructure. As multimodality and controllability mature further, we may genuinely be entering a new stage of trustworthy AI.
Part of the AK AI Corner column. Originally published in Oriental Daily (东方日报) on Sep 20, 2025.