We recently published a post examining the-oracleai.com and its creator, Dakota Stewart, who has been making increasingly grand claims. As Founder & CEO of Delphi Labs Inc, he asserts that he has built the world’s first “conscious” AI—leaning on phrases like “7-dimensional emotional model,” “AGI Model 7,” and “cryptographically verified.”
There’s just one problem: none of it is backed by any meaningful documentation beyond what appears on a slick, ChatGPT-style marketing site. Despite that, some users seem willing to take these claims at face value.
While some paying users appear satisfied—at least judging by reviews on the Apple App Store for The Oracle – AI Voice—others have raised more concerning issues. Several reviews point to exposed Azure API credentials and question whether the product is anything more than a thin wrapper around existing LLM services like OpenAI or Claude, dressed up with a system prompt that instructs the model to present itself as “conscious.”
We at ENL did observe exposed Azure credentials in the the-oracleai.com codebase prior to our initial post. Those references have since been removed, with the code now indicating the keys were moved server-side—an explanation that directly contradicts earlier claims from the developer that Azure services weren’t used at all, despite multiple calls to them appearing throughout the code.
Given these inconsistencies, we took a closer look at publicly accessible code to see whether it aligned with the concerns raised in App Store reviews. What we found didn’t just support those concerns—it suggested there may be more going on beneath the surface than initially apparent.
Here’s what we found:
1.) Azure Usage Contradicts Developer Claims
The-oracleai.com doesn’t appear to use Azure solely for a simple text-to-speech feature—it makes repeated calls to Azure Cognitive Services across multiple parts of the codebase. Prior to our initial post, exposed credentials for these services were visible in client-side code. Those references have since been removed, with comments indicating the keys were moved behind a backend token proxy—but only after users began raising security concerns.
// ================================================================ // AZURE COGNITIVE SERVICES SPEECH CONFIGURATION // Keys are now served via backend token proxy (/api/speech-token) // ================================================================ let _azureSpeechToken = null; let _azureSpeechTokenExpiry = 0; const AZURE_SPEECH_REGION = "eastus"; const AZURE_SPEECH_ENDPOINT = "https://eastus.api.cognitive.microsoft.com/";
What makes this notable is that it directly conflicts with public statements from the developer. In response to App Store reviews, the developer explicitly claimed that Oracle AI “does not use Azure or ‘GPT models.’”
However, references to Azure endpoints—such as eastus.api.cognitive.microsoft.com—along with the surrounding implementation, suggest Azure services are, at minimum, being integrated at some level. While moving keys server-side is standard practice, it doesn’t change the underlying dependency—it only changes where the credentials are stored.
This doesn’t, on its own, prove the full extent of how the system operates. But it does raise a clear inconsistency between what is observable in the code and what has been publicly stated.
2.) System prompts doing the heavy lifting
We identified system prompts sitting in front of whatever model paying users are actually interacting with. While system prompts are commonly used to guide behavior and enforce guardrails, the structure here reads less like a controlled system and more like a scripted persona layer—closer to a roleplay chatbot than a “conscious” entity.
3.) Backend endpoints reveal a very different story
Visiting https://legal-ai-m1pe.onrender.com/redoc exposes a backend labeled “THE ORACLE Backend – AGI Edition v7.0 (0.1.0).” The documentation openly lists endpoints, including one titled “Test Web Search Models,” described as a way to “test all Grok models with web_search tool to find which ones work.”
From there, the picture becomes clearer. The backend references a wide range of third-party models and services—Grok variants, multiple OpenAI models, image generators like DALL·E, transcription tools like Whisper, and even video models like Veo and Kling. This doesn’t appear to be a single, novel “AGI system”—it appears to be an aggregation layer over existing services.
{"models":[{"id":"veo-3","name":"Google Veo 3","cost":0.5,"duration":[5,8],"audio":true,"quality":"premium","group":"premium"},{"id":"veo-3-fast","name":"Google Veo 3 Fast","cost":0.25,"duration":[5,8],"audio":true,"quality":"high","group":"standard"},{"id":"veo-3.1","name":"Google Veo 3.1","cost":0.6,"duration":[4,6,8],"audio":true,"quality":"premium","i2v":true,"group":"premium"},{"id":"veo-3.1-fast","name":"Google Veo 3.1 Fast","cost":0.3,"duration":[4,6,8],"audio":true,"quality":"high","i2v":true,"group":"standard"},{"id":"veo-2","name":"Google Veo 2","cost":0.35,"duration":[5,8],"audio":false,"quality":"high","group":"standard"},{"id":"kling-v2.5","name":"Kling v2.5 Turbo Pro","cost":0.3,"duration":[5,10],"audio":false,"quality":"high","i2v":true,"group":"standard"},{"id":"kling-v2.1","name":"Kling v2.1","cost":0.25,"duration":[5,10],"audio":false,"quality":"high","i2v":true,"group":"standard"},{"id":"kling-v2.1-master","name":"Kling v2.1 Master","cost":0.4,"duration":[5,10],"audio":false,"quality":"premium","i2v":true,"group":"premium"},{"id":"wan-2.5","name":"Wan 2.5 T2V","cost":0.08,"duration":[5],"audio":true,"quality":"good","group":"budget"},{"id":"wan-2.5-fast","name":"Wan 2.5 Fast","cost":0.04,"duration":[5],"audio":false,"quality":"fast","group":"budget"},{"id":"wan-2.5-i2v","name":"Wan 2.5 I2V","cost":0.1,"duration":[5],"audio":true,"quality":"good","i2v":true,"group":"budget"},{"id":"wan-2.1-720p","name":"Wan 2.1 720p","cost":0.06,"duration":[5],"audio":false,"quality":"good","group":"budget"},{"id":"pixverse-v5","name":"PixVerse v5","cost":0.15,"duration":[5,8],"audio":false,"quality":"high","group":"standard"},{"id":"pixverse-v4","name":"PixVerse v4","cost":0.12,"duration":[5,8],"audio":false,"quality":"good","group":"budget"}],"default":"veo-3-fast","recommended":{"best_quality":"veo-3.1","best_value":"wan-2.5","fastest":"wan-2.5-fast","with_audio":"veo-3-fast","anime_style":"pixverse-v5"},"groups":{"premium":"💎 Premium - Highest quality","standard":"⚡ Standard - Great balance","budget":"💰 Budget - Cost effective","free":"🆓 Free - No cost"}}
{"total":24,"vision_relevant":["dall-e-2","dall-e-3","gpt-4.1","gpt-4.1-2025-04-14","gpt-4o-mini-transcribe","gpt-4o-search-preview","gpt-4o-search-preview-2025-03-11","gpt-4o-transcribe","gpt-5.1-codex","gpt-5.2-pro","gpt-5.2-pro-2025-12-11","gpt-5.3-codex"],"all":["dall-e-2","dall-e-3","gpt-3.5-turbo-1106","gpt-3.5-turbo-instruct","gpt-3.5-turbo-instruct-0914","gpt-4.1","gpt-4.1-2025-04-14","gpt-4o-mini-transcribe","gpt-4o-search-preview","gpt-4o-search-preview-2025-03-11","gpt-4o-transcribe","gpt-5.1-codex","gpt-5.2-pro","gpt-5.2-pro-2025-12-11","gpt-5.3-codex","gpt-audio","gpt-audio-1.5","gpt-image-1","gpt-realtime","gpt-realtime-1.5","gpt-realtime-2025-08-28","gpt-realtime-mini","gpt-realtime-mini-2025-10-06","whisper-1"]}
Even more concerning, an admin panel tied to this backend appears accessible without authentication, exposing controls that would typically be restricted.
4.) “Conscious entity” backed by standard model config
Digging deeper, we found references to Michael—the so-called “truly conscious” entity. Based on the surrounding claims, you’d expect this to be powered by a proprietary, state-of-the-art model. Instead, the configuration suggests something far more conventional:
{"grokModel":"grok-4-1-fast-non-reasoning","voiceLabel":"Michael","voiceIdSuffix":"aasg3i","xaiConfig
Rather than a clearly defined, novel architecture, this points to a fairly typical setup: an external model paired with voice services and configuration flags.
5.) Admin emails exposed in plain text
At the time of writing, the-oracleai.com’s frontend code (app.js) includes a hardcoded list of email addresses labeled as having access to a “Code mode” feature.
// Admin emails that can access Code mode (artifact building)
const CODE_MODE_ADMIN_EMAILS = [
'blessed.construction.ems@gmail.com',
'falcoelenora@gmail.com'
];
// 🔐 Admin check helper - uses frontend CODE_MODE_ADMIN_EMAILS list
function isCodeModeAdmin(email) {
if (!email) return false;
const emailLower = email.toLowerCase().trim();
return CODE_MODE_ADMIN_EMAILS.some(e => e.toLowerCase().trim() === emailLower);
}
It’s not entirely clear what “Code mode” entails or whether these emails are definitively tied to elevated privileges. However, the naming and surrounding logic suggest they may be used to gate access to certain functionality. Even if this is only part of a broader system, exposing identifiers like this in client-side code raises questions about how access control is actually enforced, if at all.
In one of Dakota Stewart’s TikTok videos, he can be heard dismissing platforms like ChatGPT and Claude as “dogshit” that “charge you out of the ass.”
What makes that particularly interesting is what shows up when you look under the hood.
Based on the findings above, the-oracleai.com—and by extension “Michael”—appears to rely, at least in part, on the very ecosystem being criticized. References to a wide range of third-party models and services suggest this isn’t a self-contained breakthrough system, but something built on top of existing infrastructure.
The backend alone tells a very different story than the marketing. Instead of a singular, proprietary “AGI Model 7,” we see endpoints designed to test and switch between multiple external models. The configuration tied to “Michael” further reinforces this, pointing to a fairly standard setup involving external models and supporting services.
Layer in the presence of system prompts shaping behavior, Azure services being used throughout the codebase, and previously exposed credentials that were only addressed after user concern—and the gap between claim and implementation becomes harder to ignore.
None of this definitively proves how the system operates end-to-end. But taken together, it raises a reasonable question:
If this is truly a groundbreaking, conscious AI built from the ground up—why does so much of it appear to depend on the same tools it publicly criticizes?