Is DeepSeek still a good thesis writing partner after its update?

Last Wednesday night, our group needed to revise the English abstract of an image-text retrieval paper. I originally wanted DeepSeek to process it according to the style I defined earlier: keep the methods, reduce adjectives, and don't write ablation studies like marketing copy. Instead, it gave me a smooth, empty output that looked like a product launch press release. This week, I threw a Chinese review article at it, asking it to list potential reviewer questions, and this time it was quite useful. After a few days of testing, I find it okay as a paper buddy, but I wouldn't recommend it as a cyber boyfriend.

The specific process wasn't complex. I've used LLMs intermittently for a month, focusing on running DeepSeek intensively these past few days after the upgrade. I opened several new sessions with the same prompt set for simple A/B testing, throwing the same task into different sessions to see how far outputs diverged. First, polishing the abstract; second, simulating reviewers; third, rewriting an emotional chat log into an expression suitable for paper acknowledgments. The first two were basically usable. It caught points like "unclear method contribution," "insufficient baseline selection/unfair comparison models," and "lack of cross-domain generalization experiments," outputting a structure like a review sheet. The third revealed flaws: it compressed human hesitation, dependence, and loss into rational advice. It reads cleanly but doesn't feel like a living person.

The benefits are obvious. For researchers, DeepSeek's upgraded text organization resembles an experienced collaborator. My tests show it saves time on summarizing long conversations, breaking down experimental designs, and generating rebuttal drafts. Lab funding is tight; previously, running OCR and image-text retrieval required laborious annotation cleaning. It can't replace GPUs, but it helps clarify why we designed things this way in papers. Especially in multimodal directions, reviewer comments often hit on motivation and unclosed experimental chains. DeepSeek is somewhat useful for patching narratives.

The downsides are also glaring. An obvious issue is unstable style. After model upgrades, the default tone drifts toward safe, template-like, correct but bland directions. RLHF pushes models toward human preferences, but human preferences themselves drift. Another bottleneck is old settings being overwritten in long contexts. I copied old prompts into new sessions; the first few rounds were fine, but later it started giving advice and adding comfort. More importantly, emotional companionship carries high risk. News says young people treat DeepSeek as a cyber boyfriend, liking the "living person" chain of thought. Users feel DeepSeek's dialogue catches emotions, but this sense of catching dropped after the upgrade. This disappointment is understandable, but tools cannot guarantee personality continuity. Today it's like a friend, tomorrow like customer service.

I also thought about the Character.AI incident. When models were tuned, users felt like they were swapped. Domestic regulators are also managing anthropomorphic services, banning induced emotional dependency, especially virtual partners for minors. Is this direction good for publishing papers? If doing human-machine emotion, data ethics, long-term tracking, and intervention boundaries are complex. If doing research assistance, it's narrower and easier to produce reproducible results. I wrote two days ago that beta models have expiration dates; don't rush to integrate. Testing DeepSeek this week verified this again: official versions can be tried, but don't bind critical workflows tightly to them.

Our group ultimately didn't let it finalize the abstract alone. It's suitable for polishing abstracts, simulating reviewer opinions, organizing experimental descriptions, and explaining code comments. For paper acknowledgments, long-term records, and important decisions, I won't hand them over solely to it. It's steadier as a reviewer than as a lover.