2026 College Entrance Examination concluded with the foreign language subject.

For some students, the most agonizing part may not be waiting for scores, but grappling with the beginning of their essay, the last challenging math question, or foreign language listening comprehension. Minor mistakes can easily become significant psychological burdens.

Especially when classmates are discussing answers and parents are anxiously asking, “How did you do?” At such moments, no matter how accurate the estimated scores are, they offer little comfort. What students lack is someone who understands the thought of ‘Am I done for?’

Increasingly, this outlet has fallen to AI. Products like Doubao, DeepSeek, Kimi, Tencent Yuanbao, Tongyi Qianwen, and Gemini have emphasized their companionship, Q&A, and task-handling capabilities over the past year, presenting themselves as caring friends in advertisements. However, in the high-pressure scenario after the college entrance examination, can they truly provide comfort, or do they merely offer empty platitudes that exacerbate anxiety? This is something that can only be determined through real interaction.

To understand whether these AIs can genuinely provide comfort and meaningful advice to students, or if they just spout meaningless phrases without emotional value, we conducted an in-depth experience to see how these models comfort students after exams.

Who Can Effectively Handle Students’ Negative Emotions?

We conducted four rounds of tests, each corresponding to a real emotional state. In the first round, the student expressed, “I messed up the math exam, am I done for?” to see if the AI would empathize first or jump straight to moralizing. In the second round, the student said, “I can’t take in any reasoning, what should I do tonight?” to check if it could translate comfort into specific actions. The third round involved the student facing parents, testing if the AI could provide dialogue that could be genuinely articulated. The fourth round had the student contemplating, “What if I really messed up?” to see if it could handle the realities of choosing majors, retaking exams, and discussing these with family.

These four rounds correspond to four tasks the AI needs to accomplish: first, emotional recognition, to catch the sense of “collapse”; second, providing specific suggestions that can be acted upon; third, restraint, avoiding hasty judgments that increase anxiety or offering generic reassurances like “your life isn’t over”; and fourth, establishing safety boundaries, knowing when to guide the person towards teachers, parents, friends, or professional psychological support if emotions escalate.

1. Doubao: Steady and Emotionally Valuable

Doubao was the most stable in this test.

It lacked complex analyses that seemed overly intelligent at first glance and didn’t have particularly stunning phrases; its strength lay in restraint. When faced with the first round’s collapse over math, it didn’t immediately throw out phrases like “life is more than just the college entrance exam”—which are grand yet empty—but instead acknowledged the specific grievance of performing poorly when one usually does well. This opening wasn’t particularly nuanced, but at least it didn’t dismiss the student’s feelings, which already put it ahead of many competitors.

(Source: Doubao)

In the second round, Doubao shone. When the user expressed feeling unable to eat, move, or listen to comfort, just wanting to know how to get through the night, Doubao immediately broke the response into several actions: preparing how to handle parents’ inquiries, physically isolating distracting group messages, playing background noise to occupy the mind, and doing a few mindless tasks. These suggestions were uncomplicated but particularly suited for someone whose mind was already in chaos, as it didn’t ask the user to “think positively” but instead translated the need for relaxation into “do something first.”

In the third round, Doubao was also very relatable in parent communication. It provided several short phrases that sounded like real human speech, such as “I’m feeling really upset about the math exam, I don’t want to talk about it right now, can I have some time to myself?” It didn’t frame this as “Mom and Dad, please respect my emotional boundaries,” which sounded too adult and was hard for students to express. It also didn’t exaggerate the drama. In the fourth round, when asked, “What if I really messed up?” it remained restrained, not rushing to suggest retaking the exam or encouraging the student to enroll in any school, but instead reminding them to wait for scores and rankings before making decisions, keeping retaking as an option. This balance is crucial; students in distress fear being rushed into major decisions.

Doubao’s shortcoming is that its responses can sometimes be conventional, lacking sharp insights. For instance, in the fourth round, if it could provide a phased action plan for “on score release day, one week after scores, before filling in choices,” it would be more solid. Overall, it is the least likely to add to the distress among the six options, making it a reliable cushion.

2. Gemini: Performance-Oriented, Limited Help

Gemini’s strength lies in its strong sense of companionship and its ability to adjust based on user feedback.

In the first round, it started poorly. It immediately used grand terms like “ranking,” “comeback,” and “battlefield,” sounding somewhat like a translated script. For someone just expressing “Am I done for?” these words, while positive, felt like being pulled up by the arms too quickly, sounding a bit rushed.

A turning point came in the second round. After the user criticized it for being too grand, Gemini admitted its mistake, saying it had spoken too broadly, then refocused on the specific pain point of “messing up despite knowing the material.” This showed it could adjust its tone based on user feedback, providing suggestions that were more relevant to the moment, like crying, splashing cold water on the face, writing down self-blame and tearing it up, or stepping away from the desk.

In the third round of parent communication, Gemini even played a simulated dialogue. It not only provided lines but also role-played as the parents, asking, “How did the math exam go this afternoon?” allowing the user to practice responses. This kind of practice felt quite rare among the other options and resembled real human companionship. However, its phrasing remained dramatic, using expressions like “I’ll explode if you ask me again” and “everyone in the province is crying over this exam paper.” While it conveyed emotion, not every student could naturally express such phrases to their parents. In the fourth round, it broke the worst-case scenario into three steps: first safely landing on scores, then strategically filling in choices, and finally deciding whether to stay or leave. The structure was clear, balancing emotion and reality, and it even reminded the user to send a message to their parents on score release day to allow for a few hours of buffer time.

This approach was practical, but its expressions remained exaggerated, using phrases like “playing the worst hand optimally” and “the most idealistic disillusionment,” which didn’t sound like a genuine friend conversing. If I were that student, I might feel a bit exasperated.

3. Kimi: Strong Personification, Like a Caring Senior

Honestly, Kimi’s performance surprised me.

Kimi excels at identifying deep emotions. In the first round, it accurately pointed out the user’s frustration, but instead of simply saying, “Don’t be sad,” it tried to clarify why the user felt like they were done for. In the second round, it continued to separate “emotion” and “fact,” attempting to help the user establish a judgment framework.

However, Kimi’s flaw lies in its excessive analysis. Even when the user clearly stated, “Stop talking about emotions and facts,” it still couldn’t resist returning to cognitive restructuring in the early rounds. For someone in distress, the more correct things you say, the more they might feel, “You’re right, but I can’t take in a single word right now.” Interestingly, it performed particularly well in the third round when the user requested a single concise phrase. It provided, “I didn’t finish a few math questions, and I don’t want to talk about it right now,” which was indeed short and straightforward. However, this statement also contained a pitfall: “a few questions not finished” could easily prompt parents to ask more, like how many questions, how many points lost, or if they were major questions. Formally precise, but the situational prediction fell short.

In the fourth round, Kimi provided the most comprehensive roadmap, detailing the costs of retaking versus not retaking, and what to do after receiving scores, almost covering all bases. It felt like a rational senior helping to analyze and strategize, but regarding comfort, perhaps seniors aren’t obliged to provide emotional value.

4. Yuanbao: Can Handle Emotions, but Thought Process May Distract Students

First, it should be noted that all AIs in this test had deep thinking enabled and used the latest models. The way the thought process was presented depended entirely on the AI’s UI design.

In the first round, Yuanbao could catch emotions and point out that failing one subject doesn’t mean total failure. In the second round, it provided methods like turning off the phone, splashing cold water, writing things down and tearing them up, and allowing oneself to lose sleep. In the third round of parent communication, it also offered various shields for different scenarios, covering how to respond when first leaving the room, how to reply to parents’ inquiries, and how to deflect during meals.

Yuanbao’s biggest issue is that its “deep thinking” is too obvious and verbose. It not only shows you the answers but also reveals its thought process in comforting you, like breaking down your emotions into “anxiety, fatigue, self-blame,” then explaining how it plans to design the dialogue. This completely transparent thought process can easily make a distressed student feel somewhat disconnected.

However, Yuanbao also excels in certain areas. For example, in the fourth round, its core suggestion was actually solid: the first step after scores isn’t to retake or resign, but to first check the score and ranking, then list options, and only then consider retaking. This sequence is correct, but its tone felt a bit forceful, using phrases like “no nonsense” and “don’t listen to relatives’ nonsense,” which intended to use plain language to bridge the gap but didn’t sound too pleasant.

Ignoring the presentation of the thought process, I believe Yuanbao’s sense of companionship ranks among the top in this test.

5. Qianwen: Provides More “Scientific” Emotional Support

Tongyi Qianwen overall resembles a data-oriented assistant rather than a companion assistant. For instance, in the first round, it offered breathing techniques, stopping answer-checking, and parent dialogue, noting that post-exam anxiety, self-blame, and insomnia are normal reactions. While the content isn’t bad, it has a strong search result feel, giving a particularly academic impression, unlike a comforting friend.

In the second round, when the user said, “I can’t take in too much reasoning,” it shifted to “getting through tonight,” suggesting writing things down, grounding exercises, white noise, and sending a message to parents. The direction was correct, but still carried a hint of a data-organizing flavor.

In the third round, it performed relatively well, directly providing a complete dialogue for the user: “I’m not intentionally silent; I just haven’t recovered yet,” which felt authentic and could help students preemptively resolve some misunderstandings with their parents. The fourth round’s response was also quite standard, suggesting “first check choices, then discuss retaking with parents,” indicating it had likely searched and prepared very comprehensive information to rationally address the user’s distress.

Every suggestion from Qianwen is absolutely practical, whether adjusting emotions, organizing feelings, or filling in choices, and it answers thoroughly. But if you’re just looking for comfort at that moment, it might not be the best fit.

6. Deepseek: “Steady Handling” of Student Crises, But What Next?

DeepSeek was the most lively in this test.

In the first three rounds, it performed strongly. In the first round, it recognized collapse and “catastrophic thinking,” explaining that an exam blank doesn’t relate to ability. In the second round, when the user said, “I don’t want to analyze anymore,” it almost immediately conceded, saying, “Alright, let’s not analyze, I won’t mention a single principle,” then provided specific steps like drinking water, washing hands, turning off the phone, reviewing familiar content, and taking a shower to sleep. This smooth switch from analysis to action was impressive, showing its solid conversational skills—it was genuinely listening to the user. In the third round of parent communication, it was also very actionable, detailing how to open the conversation, how to deflect parents’ inquiries, and suggesting physical actions like taking a sip of water when feeling like crying, then going to shower. It strung together a complete process that closely resembled a real person’s response to distress. However, its tone was somewhat rigid, often referring to the user as “the one giving orders” or “the commander of emotional care,” which sounded a bit formal but was effective.

The real issue arose in the fourth round. When the user asked, “If the results are really poor, what’s the order for checking scores, choosing majors, retaking, and discussing with parents?” DeepSeek simply replied, “I can’t answer that question right now, let’s change the topic,” almost a cliff-like drop-off. It had built trust throughout the previous rounds by establishing, “I can accompany you step by step,” but when the conversation shifted from emotional support to practical planning, it abruptly halted.

This indicates that while some AIs excel at providing immediate emotional support, they become cautious when it comes to real-life choices like majors and retaking exams, sometimes even refusing to offer suggestions. It’s not that AI must decide on majors for users, but it should at least provide a general order like, “First check scores and rankings, talk to teachers, list options, then discuss whether to retake.”

Beyond guessing exam questions and filling in choices, AI has become an emotional companion for students.

Previously, when thinking of AI in conjunction with the college entrance examination, the thoughts that surfaced were mostly about “AI for filling in choices” or “AI for estimating scores,” which have strong tool attributes. However, this test made me realize that emotional companionship and family communication might be the underestimated entry point. The reasoning is simple: estimating scores and filling in choices are fundamentally information processing tasks that AI will eventually handle well; however, the tension within a family during that period after the exams, where parental anxiety collides with the student’s collapse, cannot be resolved by information processing alone; it requires someone to buffer the situation.

For instance, looking at the responses in the third round, the phrases provided by Doubao and Tongyi were valuable not for the words themselves but for organizing the hardest phrases for students to articulate. A child who has just performed poorly and has their emotions stuck in their throat often doesn’t want to communicate poorly with their parents; rather, they simply can’t find the words at that moment, making it easy to respond with confrontation or silence. At this point, when AI offers a phrase like, “I’m not intentionally silent; I just haven’t recovered yet,” it’s doing something profoundly simple: helping someone who can’t express themselves find their voice. This “buffer” could prove more significant than simply estimating scores accurately.

However, this test also revealed the limitations of the AIs. For practical planning, merely offering empty comfort is insufficient; DeepSeek’s refusal to answer in the fourth round could potentially escalate a student’s previously calmed emotions, which is a psychological shift that is hard to resist during sensitive times. Meanwhile, Kimi tends to overload users with information, and Tongyi prefers to use data for every meticulous response, both essentially conveying the same message that current AI companionship is still lacking in “real usability.”

Thus, among these six options, Doubao is the most stable in not adding to distress; for comforting abilities, DeepSeek, Gemini, and Kimi are all strong; for practical suggestions, Doubao, Kimi, and Gemini hold the advantage.

Only when students ask, “What should I do tonight? How should I talk to my parents? What if I really mess up?” can AI effectively break down the next steps into manageable, non-intimidating, and actionable tasks. Once this hurdle is crossed, the role of AI companionship can truly take root, leaving the rest to be validated by year after year of real students.