Pada si Iroyin
Awọn iroyinAI Understanding finifini

Iwadii Wa Awọn iwiregbe AI Kukuru Igbagbọ Dinku ninu Awọn Idite Tuntun

Awọn adanwo AMẸRIKA meji ti rii pe awọn ibaraẹnisọrọ Gemini ti o ṣe deede dinku igbagbọ awọn olukopa ninu awọn imọ-jinlẹ tuntun ti o ṣẹda diẹ sii ju awọn ibaraẹnisọrọ ti ko ni ibatan ati nigbagbogbo diẹ sii ju awọn iwe otitọ aimi lọ.

5 min readRead the primary source
Iwe aṣẹ orisun akọkọOrisun ti o gbasilẹ
Olutẹwe
Costello and colleagues' research paper on arXiv
Orisun ọna asopọ
arxiv.orghttps://arxiv.org/abs/2608.06151
Orisun iru
Iwe akọkọ - ikede osise, iwe, iforukọsilẹ, tabi oju-iwe ẹgbẹ akọkọ ti a ka taara.
AtokọLoye eyi ni iṣẹju 60

Bẹrẹ nibi

Awọn ofin bọtini

Iyasọtọ
Iṣẹ-ṣiṣe nibiti awoṣe kan ti n fi igbewọle si ọkan tabi diẹ ẹ sii awọn ẹka ti a ti ni asọye.
Ṣe idanwo fun ara rẹAI Ethics adanwo

Kini o ṣẹlẹ

A Carnegie Mellon, MIT, ati Cornell iwadi egbe Pipa meji laileto irú-ẹrọ lori August 6 igbeyewo boya kukuru AI awọn ibaraẹnisọrọ le koju rikisi igbagbo nigba ti mon nipa pataki iṣẹlẹ si tun nyoju.

Awọn oniwadi ṣe atupale awọn agbalagba AMẸRIKA 472 ti o ṣafihan awọn iwo iditẹ lẹhin igbiyanju Keje 2024 lati pa Donald Trump ati 1,035 lẹhin ipaniyan Oṣu Kẹsan 2025 ti Charlie Kirk. Awọn olukopa ni a yan laileto si ibaraẹnisọrọ debunking ti a ṣe deede, ibaraẹnisọrọ AI ti ko ni ibatan, tabi atokọ aimi ti awọn otitọ asiko.

Awọn ẹgbẹ ijiroro naa pari o kere ju awọn paṣipaarọ marun pẹlu Gemini 1.5 Pro ni idanwo akọkọ ati Gemini 2.5 Pro ni keji. Awọn awoṣe mejeeji gba ipilẹ otitọ kan ti o ni iyasọtọ ti o yapa alaye ti a fọwọsi, awọn ẹtọ ti a sọ di mimọ, ati awọn ibeere ti ko yanju; awoṣe keji tun le wa oju opo wẹẹbu nikan lati rii daju alaye otitọ.

Lori iwọn 0-si-100 ti igbagbọ ninu imọ-ọrọ asọye ti alabaṣe kọọkan, awọn ijiroro ti o ni ibamu ṣe agbejade awọn idinku ti awọn aaye 6.95 ni ibatan si iṣakoso iwiregbe ti ko ni ibatan ni idanwo akọkọ ati awọn aaye 7.56 ni keji. Awọn iyatọ lati iwe otitọ aimi jẹ 5.60 ati awọn aaye 6.56. Iwe naa ṣe ijabọ awọn ipa idiwọn ti aijọju 0.32 si 0.38 fun awọn afiwera yẹn.

The two case studies were designed around moments when the factual record was still developing. The first followed the July 2024 attempt to assassinate Donald Trump; the second followed the September 2025 assassination of Charlie Kirk. Participants were screened for conspiratorial or uncertain interpretations, then assigned to a conversation that addressed their own stated theory, an unrelated AI chat, or a static contemporaneous fact sheet. The second study was preregistered, while the first was not, so the replication is informative but not equivalent to two independent confirmatory trials.

Awọn alaye orisun: Costello and colleagues' research paper on arXiv

Kini idi ti o ṣe pataki

Abajade jẹ imọran pe eto ibaraẹnisọrọ le ṣe diẹ sii ju atunṣe atunṣe kan lọ: o le ṣe atunṣe awọn otitọ, awọn ibeere, ati aidaniloju si idi pataki ti eniyan fi funni fun igbagbọ kan.

Iyatọ yẹn wulo lakoko awọn iṣẹlẹ gbigbe ni iyara, nigbati ẹri ti o rii daju ko pe ati pe iwe otitọ jeneriki le ma koju ẹtọ ti eniyan rii ni idaniloju. Atunyẹwo ilana iwe naa rii pe awoṣe naa da diẹ sii lori awọn orisun igbẹkẹle, awọn ibeere ṣiṣi, ati iṣọra nigbati a ko mọ diẹ, lẹhinna lo ẹri taara diẹ sii nigbati iṣẹlẹ keji ni igbasilẹ otitọ ti o tobi julọ.

Awọn onkọwe tun jabo ẹri ti o lopin ti itusilẹ nigbamii. Oṣu meji lẹhin idanwo akọkọ, awọn olukopa ti a yàn si ijiroro ko ṣeeṣe ju awọn iṣakoso akojọpọ lati fọwọsi awọn itumọ iditẹ meji ti iṣẹlẹ ti o tẹle. Ni atẹle keji, iṣẹ iyansilẹ itọju taara ko dinku awọn igbagbọ ni pataki nipa ibon yiyan nigbamii, botilẹjẹpe itupalẹ itẹramọṣẹ iṣaaju ti o yatọ ati iwọn iditẹ gbooro daba awọn ipa gbigbe-lori kekere.

The design shows why the interaction may outperform a static correction without proving that persuasion is always desirable. Gemini 1.5 Pro in the first study and Gemini 2.5 Pro in the second received researcher-curated fact bases separating confirmed information, debunked claims, and unresolved questions. The model could ask about the participant's specific reasoning and choose whether to answer directly, cite credible evidence, or preserve uncertainty. That adaptability is useful for education, but it also gives the system discretion over which claims to challenge and which sources to foreground.

Iwadi naa ko ṣe idalare ni gbigbe awọn botilẹnti ipalọlọ sinu awọn ibaraẹnisọrọ gbangba. Eto ti a fun ni aṣẹ lati yi awọn igbagbọ pada ni agbara dani, ati awọn onkọwe ṣe akiyesi pe awọn ilana ti o jọra le tun mu igbagbọ pọ si awọn ẹtọ eke. Eyikeyi iṣẹ gidi yoo nilo iwe-aṣẹ ti o han gbangba, ilẹ iṣẹlẹ ti o gbẹkẹle, awọn aabo lodi si ibi-afẹde iṣelu, ati awọn idanwo ominira ti deede ati awọn ipa airotẹlẹ.

Interactive Mechanism

Ibaraẹnisọrọ Mechanism: Bii O Ṣe Nṣiṣẹ Lootọ

Ṣawari imọ-ẹrọ abẹlẹ lẹhin idagbasoke yii ni ibaraenisọrọ.

Thinking Budget (Test-Time Tokens):1,024 tokens
Complex Accuracy79%Math & Code Logic
Latency3.2sTime to first full output
Inference Cost$0.0092Per query estimated
Reasoning StyleStep VerificationInternal chain depth
Active Thinking Trace:
1Deconstruct user problem into formal constraints
2Propose candidate hypotheses & step-by-step calculation
3Self-correction: Backtrack and refute subtle edge cases
4Exhaustive consistency check & final output synthesis
Core takeaway: Test-time compute fundamentally changes AI economics. Instead of only scaling during pre-training, giving reasoning models more tokens at inference time allows them to systematically solve PhD-level STEM problems.
Ibanisọrọ Erongba Ṣayẹwo+10 Points
AI Ethics Quiz

Impossibility results in algorithmic fairness (e.g. Kleinberg et al., Chouldechova) show what?

Kini lati wo tókàn

Ṣọra fun atunyẹwo ẹlẹgbẹ, ẹda ti o kọja awọn rogbodiyan iṣelu AMẸRIKA meji, ati ẹri aaye ti n fihan boya eniyan yan lati lo iru ohun elo laisi gbigba sinu ikẹkọ kan.

Eyi jẹ iwe atẹjade tuntun ti a ṣe ni ayika awọn iwadii ọran meji. Idanwo akọkọ ko ni iforukọsilẹ tẹlẹ, ekeji jẹ, ati pe awọn mejeeji ṣe iwọn awọn igbagbọ ti ara ẹni ti o royin lẹsẹkẹsẹ lẹhin ibaraenisepo kukuru lori ayelujara kuku ju ihuwasi pinpin agbaye gidi, idibo, tabi igbẹkẹle igba pipẹ.

Awọn apẹẹrẹ ti a ṣe atupale pẹlu awọn olukopa nikan ti awọn idahun ṣiṣi jẹ ipin bi iditẹ tabi aidaniloju nipasẹ GPT-4o, botilẹjẹpe awọn onkọwe jabo awọn ilana ti o jọra labẹ awọn ofin isọdi omiiran. Awọn ijinlẹ mejeeji gba awọn agbalagba AMẸRIKA ṣiṣẹ nipasẹ CloudResearch, nitorinaa awọn awari le ma gbe lọ si awọn orilẹ-ede miiran, awọn ede, awọn iṣẹlẹ, tabi awọn olugbe.

Awọn awoṣe ko gbarale imọ ti a ko ṣe iranlọwọ: awọn oniwadi pese awọn ipilẹ otitọ ti o ṣajọpọ ni iṣọra ati awọn ilana itusilẹ gbangba. Iṣẹ iwaju yẹ ki o ṣe idanwo tani o ṣetọju awọn ododo wọnyẹn labẹ titẹ akoko ipari, bawo ni a ṣe ṣe atunṣe awọn aṣiṣe, boya awọn oju-ọna atako ni a tọju nigbagbogbo, ati nigbati eto kan yẹ ki o tọju aidaniloju dipo igbiyanju lati yi pada.

There is also a measurement distinction between changing a stated belief and improving a person's understanding. The outcomes were self-reported ratings collected after short online conversations, not observed sharing behavior, source checking, voting, or decisions made during a live crisis. The paper's follow-ups provide early evidence about persistence, but the mixed results mean a later study should pre-register long-term outcomes, track attrition, and test whether participants can explain the evidence themselves rather than simply repeating the model's conclusion.

Replication should vary the source of the factual record as well as the population. These experiments supplied researchers' own fact bases and recruited US adults through CloudResearch, which makes the setup unusually controlled but leaves open questions about multilingual events, lower-connectivity settings, and crises in which official information is delayed or disputed. A responsible field trial would make the model's authorship visible, let participants decline the intervention, log which evidence was shown, and use independent adjudicators to check whether the dialogue corrected a misconception without introducing a new one. It should measure trust, emotional response, and willingness to share claims alongside belief scores.

Awọn itọsọna ti o jọmọ & awọn ibeere

Ìlànà Ìwà AIAI AaboKini AI?Ṣe idanwo ohun ti o mọ — gbiyanju idanwo AI ọfẹ kanWa ọrọ AI kan ninu iwe-itumọ wa
Ṣe eyi wulo?