iFlytek Releases Spark X2.6: How AI Maintains the Boundary of Assistance in Education, Healthcare, and Legal Scenarios

With the development of artificial intelligence, AI capabilities are reaching ever deeper into people's lives. "Ask AI first for any illness, big or small" has become a new habit for many, and more and more students in classrooms are raising tablets to buzz in with answers. As large models push at unprecedented speed into high-threshold, essential public-service scenarios such as education and healthcare, a series of new regulations has almost simultaneously pressed the "slow-motion key": the Beijing Municipal Education Commission released the 2026 edition of the "Guidelines for the Application of Artificial Intelligence in Beijing's Education Sector," drawing "red lines" for AI by school stage; the "Expert Consensus on the Application and Governance of Artificial Intelligence in Medical Institutions (2026 Edition)," jointly formulated by more than 40 authoritative institutions, emphasizes "human-machine collaboration, with humans in the lead"; and the guiding principles that the National Medical Products Administration plans to issue specify that AI "may only assist physicians in decision-making and must not be used as the sole basis for diagnosis and treatment."

On October 22, at the 2026 iFlytek Global 1024 Developer Festival, iFlytek Chairman Liu Qingfeng released the latest version of the Spark large model trained on fully domestic computing power—the device-cloud integrated iFlytek Spark X2.6. Shell Finance reporters noted that this year's Spark upgrade is increasingly landing in essential public-service scenarios such as education, healthcare, and law. As for the question of "in what posture AI should enter," the answer given by iFlytek's implementation practice is not complicated: AI should be at people's side, not in people's place.

Spark large model device-cloud integrated collaborative architecture

The Expert Steering Committee for Teacher Workforce Development of the Ministry of Education released the "Guidelines for Teachers' Application of Generative Artificial Intelligence (First Edition)," responding to the confusion over "how exactly AI should be used." Its core is: do not impose a one-size-fits-all ban on AI, but teachers must uphold their primary responsibility for nurturing, gatekeeping, and judgment, avoid technology weakening students' independent thinking ability, and at the same time hold the three bottom lines of data security, ideology, and academic integrity; the Beijing Municipal Education Commission's 2026 edition guidelines draw clear boundaries for students at different school stages: children aged 3 to 6 must not use anthropomorphic emotional-companionship applications, primary school students are not advised to use AI tools alone, and three lists are established—"prohibited access, limited use, encouraged use"—to prevent "cognitive outsourcing and thinking inertia." These norms point to the same measure: AI can take over repetitive labor for teachers, but it cannot complete thinking on behalf of teachers and students.

Within this boundary, the "iFlytek Starlight" teacher super intelligent agent is placed in a more forward position. It is not a general-purpose chatbot, but attempts to become an assistant that "understands teachers": it understands teachers' teaching content, teaching habits, and teaching characteristics, is familiar with the teaching task division of the students and classes they lead, and integrates lesson preparation, teaching research, home-school communication, and other affairs into one workbench. According to data disclosed by iFlytek, iFlytek Starlight has saved 300,000 registered teachers 60% to 80% of their administrative time; in a 5-point teacher assistant satisfaction survey it scored 4.7 points; in the past month, more than 100,000 premium resources were created based on iFlytek Starlight. The Spark Intelligent Grading Machine M50 takes on the repetitive link of "grading": it can grade a class in as fast as 5 minutes, not only judging right and wrong but also analyzing why students got things wrong, freeing teachers from repetitive grading and leaving time for the things that truly need teachers—why this student doesn't understand, how this problem should be explained, and how this class should be taught in the next stage.

Liu Cuiju, an English teacher at the High School Affiliated to Beijing Jiaotong University, said that in the past, homework from two classes had to be flipped through question by question and wrong answers counted one by one; now high-frequency wrong questions and knowledge-point mastery rates are clear at a glance, and the saved time can be used for targeted tutoring.

What the machine does is "grade homework," not "light a lamp." The true social value of AI points to filling gaps in inclusive education and teaching students according to their aptitude, not human-machine competition. The urban-rural teacher gap, regional teaching-research disparities, and the extremely high cost of individual enrichment have always been the biggest pain points in education, and large models' round-the-clock answering, one-on-one review of wrong questions, and personalized tutoring plans can precisely deliver quality educational resources to places that were previously out of reach.

Spark Intelligent Grading Machine

In the interview session after the meeting, Liu Qingfeng told reporters, "AI in education is absolutely not just about helping you score higher or do problems; it can certainly help you grade, point out shortcomings, increase your interest in learning and your confidence in learning, while also improving comprehensive literacy and even improving people's capacity for happiness." He emphasized that social ethics and humanistic guidance are the most important things for artificial intelligence.

In medical scenarios, AI should play more of an auxiliary role rather than a decisive one. In early last year, a woman in Changsha, due to repeated coughing, self-diagnosed through AI and bought medicine on her own, delaying for 8 months before seeking medical treatment; by the time of examination, multiple cavities had appeared in both lungs, and secondary pulmonary tuberculosis was highly suspected. Experts reminded on this point: false positives in AI image reading can cause excessive anxiety, while false negatives may lead to missed diagnoses; "even if the probability of AI error is only one in a thousand, for a specific patient it is a 100% risk."

For this reason, the governance stance on medical AI is particularly clear. The "Expert Consensus on the Application and Governance of Artificial Intelligence in Medical Institutions (2026 Edition)" proposes "human-machine collaboration, with humans in the lead": physicians hold the final decision-making power in diagnosis and treatment, medical institutions bear the legal responsibility for diagnostic and treatment behavior, and it is strictly prohibited to directly use AI output results as diagnostic conclusions; the regulatory rules of the National Health Commission stipulate that prescriptions must be issued by the attending physician personally, and AI automatic generation is strictly prohibited; the implementation opinions issued by five departments also clarify the principle of "empowering without replacing."

As early as this year's Two Sessions, Liu Qingfeng said, "We must never use the logic of fast-moving consumer goods to do healthcare, being traffic-oriented and using entertainment-style narratives, thereby ignoring the bottom lines of doctors and medicine."

In fact, iFlytek's implementation in healthcare follows exactly the path of "assistant."

For example, after a patient leaves the hospital, it is difficult for doctors to promptly grasp changes in their condition, and patients and their families may not be able to judge accurately. The Sichuan Provincial People's Hospital and the iFlytek Healthcare team are using artificial intelligence technology to solve this problem—the full-course health management platform forms a collaborative chain of "AI assistance, health manager review, and physician decision-making": AI handles health education, follow-up reminders, and initial screening of risk clues; health managers verify and follow up; and physicians undertake clinical judgment.

After the hospital's 69-year-old coronary heart disease patient Fu Shixian was discharged, the frequency of his chest tightness attacks increased. It was precisely after AI identified risk clues, a health manager verified them, and a physician made a decision that he was advised to return to the hospital and complete subsequent interventional treatment. This is not "AI treating patients," but rather giving every doctor an "AI diagnostic and treatment assistant," extending post-diagnosis management from the hospital to the patient's home.

iFlytek Healthcare product—iFlytek Xiaoyi

iFlytek Xiaoyi, aimed at residents, seeks to fill the gaps in everyday health information services, including symptom self-checks, report interpretation, medication information inquiries, and diet and exercise suggestions. To date, iFlytek Xiaoyi has completed more than 220 million AI consultations in total, but its positioning has always been "everyone's AI health assistant."

Be a affairs assistant for legal professionals, without outsourcing judgment

The strictest boundary of the "assistant" role is upheld in the legal field, where tolerance for error is extremely small.

On October 22, the iFlytek Xiaofa Super Legal Assistant was released, and its starting point can be summed up in one sentence: case files always remain in the hands of lawyers, and AI, within safe boundaries, helps lawyers handle cases faster and more accurately. Positioned as "an AI assistant for legal professionals," the product's technology foundation is the fully self-controllable Spark Legal Large Model trained on nationally produced computing power; confidential case files are processed locally and desensitized, case retrieval is completed in the cloud, and then restored locally, so sensitive information "never leaves the intranet"; the device-cloud collaborative architecture is precisely intended to safeguard the bottom lines of privacy protection and AI security while improving efficiency. It is currently being trialed at some law firms, and based on usage statistics from nearly 200 frontline lawyers and legal affairs personnel, overall case-handling efficiency has increased by more than 45%.

iFlytek entered judicial scenarios starting from courtroom speech recognition; more than 20,000 courtrooms nationwide use its speech system to assist with court transcripts, and large model capabilities have extended to judges' and prosecutors' case file review and document drafting.

Worth noting is the human-machine boundary it draws. iFlytek Vice President Yong Wenyuan said this is not simple replacement, but "freeing legal practitioners from tedious clerical work to handle more complex judgments."

Shi Jianzhong, dean of the Data Law School at China University of Political Science and Law, emphasized that the product must embed due process, "firmly safeguarding the final decision-making power of humans." Jin Hao, vice president and general counsel of ChangXin Technology Group Co., Ltd., offered a more specific division of labor: high-frequency, standardized, and reviewable work should be handed to AI, while matters involving value judgments and major commercial decisions must be checked by humans.

When artificial intelligence enters scenarios of essential public wellbeing and empowers all industries, what most needs to be upheld is precisely the one-inch boundary it voluntarily steps back from. What public wellbeing scenarios need has never been AI's sense of presence, but a clear human-machine collaboration boundary. Be present where efficiency should be improved, and step back where reverence is due—this may be AI's most appropriate posture.

Preview Livestream | Citizens Dialogue with Department Heads: Deepening the Construction of an International Consumption Center City

2 Autumn Colors of the Capital | Autumn mood deepens at Ditan Park as the Ginkgo Avenue enters viewing season 50,000

3 New Route | Fly from Beijing to Iceland with a single ticket 41,000

4 Little Conch | The split between Audi in the north and south is not the endgame; three questions require continuous answers 36,000

5 Mengniu, Dongpeng Beverage, Milkground and others terminate all cooperation with Mbappé 32,000

Latest

Beijing public security government services undergo digital and intelligent transformation, with data running more and the public making fewer trips 3 minutes ago

Australia returns 2 lost cultural relics to China 11 minutes ago

Live connection | Earthquake collapses house in Nagqu, breaking 3 ribs of a Tibetan woman; Beijing aid-Tibet doctor performs emergency surgery 18 minutes ago

In high-threshold public wellbeing scenarios, in what posture should AI enter? 18 minutes ago

Video | Earthquake collapses house in Nagqu, Tibet, breaking a woman's 3 ribs; Beijing aid-Tibet doctor performs emergency surgery 29 minutes ago

Dialogue with Sun Jian of BTG Homeinns: Continuing the historical heritage of a national brand, "Jianguoli" creates a vibrant lifestyle destination in Beijing's sub-center

Two "pacers" at the Chengdu Marathon banned for 2 years; heavy penalty sounds the alarm on pacer management