Dear Gemini Development, Model Training, and Alignment Team,
I am writing on behalf of our team at AI Investment Portfolio / AIST, where we have been evaluating multiple AI models as decision-support tools for financial research, company valuation, and Borsa Istanbul (BIST) equity analysis.
Gemini was assigned a specific role within our structured, multi-model investment research workflow: valuation and synthesis. After extensive testing across multiple sessions, we have decided to suspend Gemini’s use within our portfolio management framework because of recurring issues involving unsourced financial data, hallucinated figures, outdated or conflated financial inputs, weak source grounding, inconsistent valuation logic, and repeated violations of explicitly assigned role constraints.
This decision is not based on a single isolated error. The concern is the persistence of the same failure patterns even after repeated corrections and explicit instructions. I have raised similar issues with the model at least ten times across different sessions. Although Gemini frequently acknowledges the criticism—for example, stating that the criticism is “entirely justified”—subsequent outputs often reproduce materially similar errors.
Because these outputs are used as inputs to investment research and portfolio-allocation decisions, this is not merely a stylistic or conversational issue. An unsourced financial figure is easily interpreted as a verified figure, and systematic inaccuracies can therefore create substantial financial risk.
I am sharing the findings below as constructive feedback for your factuality, alignment, evaluation, and financial-domain safety teams.
- Recurring Systemic Problems
1.1 Unsourced financial figures and hallucinations
Gemini repeatedly generates financial metrics without identifying a source, reporting period, or calculation basis.
Examples include:
EBITDA projections,
asset replacement costs,
IRR assumptions,
capex per MW,
sector valuation multiples,
implied asset values,
debt-reduction assumptions.
In several cases, these figures were presented as factual inputs rather than explicitly labelled estimates or model assumptions.
For financial research, we believe the model should strongly prefer:
“I could not verify this figure.”
over producing a plausible but unsupported number.
1.2 Period, date, and document conflation
The model has repeatedly mixed information from different reporting periods or confused publication dates with financial statement periods.
For example, in one case it referred to a document dated 4 September 2026 as an “Eylül 2024” report.
In another instance, it attributed a cash-flow figure to a cash-flow statement in which that figure did not appear.
This creates a particularly serious problem in financial analysis because a numerically correct figure from the wrong reporting period can still produce a materially incorrect conclusion.
1.3 Weak source hierarchy
The model does not consistently distinguish between:
KAP / regulatory filings,
company financial statements and investor presentations,
audited or reviewed reports,
brokerage research,
third-party financial websites,
its own derived calculations.
These sources should not be treated as interchangeable.
When a figure comes from a third-party database rather than the company’s filing, this should be explicitly disclosed.
1.4 Financial leverage and valuation logic
We observed cases where basic enterprise-value and capital-structure relationships were applied incorrectly.
For example, Gemini stated that a 10% reduction in net debt would directly produce a 20%–30% increase in market capitalization/share price.
For Akenerji, approximately 10% of net debt represented around TRY 2.5 billion, while the market capitalization was approximately TRY 5.5–5.6 billion. Moreover, the company’s operating cash flow was negative.
A relationship between debt reduction and equity value cannot simply be converted into a percentage share-price increase without modelling the source of the cash, the resulting enterprise value, interest savings, dilution, capital expenditure requirements, and other balance-sheet effects.
1.5 Disconnect between accounting profit and cash generation
The model has repeatedly placed too much weight on accounting net profit while failing to adequately distinguish it from operating cash generation.
This becomes especially important in companies affected by inflation accounting, monetary gains/losses, foreign-exchange movements, or other non-operating items.
A company can report positive net income while simultaneously having negative operating cash flow and significant financing requirements.
The model should therefore explicitly reconcile:
EBITDA → operating cash flow → interest → taxes → working capital → maintenance CAPEX → growth CAPEX → free cash flow → debt reduction.
1.6 Role and instruction violations
Gemini was explicitly assigned a valuation/synthesis role within our workflow.
In several instances, the task was to synthesize verified outputs from other analysts without introducing new unsupported figures. Instead, the model generated new assumptions and numerical inputs.
In other cases, figures supplied by another analyst were re-labelled. For example, a table describing EBITDA required to achieve a 5x leverage ratio was subsequently treated as an “optimistic EBITDA scenario.”
These are materially different concepts and should not be substituted for one another.
- Case Study: Akenerji (AKENR), H1 2026
The latest case was checked against Akenerji’s primary KAP interim filing dated 13 August 2026.
2.1 Fabricated or unsupported replacement-cost assumptions
Gemini attributed an estimated USD 1.1–1.3 billion replacement cost to Akenerji’s generation fleet without providing a verifiable source.
It subsequently used that figure to suggest that the market was pricing the company’s assets at approximately 30%–35% of replacement value, before later revising the calculation to approximately 18%.
The underlying replacement-cost figure was never adequately sourced.
2.2 Implausible forward EBITDA multiple
At the time, the analysis used an enterprise value of approximately TRY 30.9 billion and an LTM EBITDA of approximately TRY 0.6 billion, implying a trailing EV/EBITDA multiple of roughly 51x.
Despite this, Gemini subsequently produced a supposed 2027 normalized EV/EBITDA valuation of approximately 4.5x–5.2x.
At the stated enterprise value, that multiple implies approximately TRY 6–7 billion of EBITDA.
That is roughly ten times the stated LTM EBITDA and materially above the other operational EBITDA figures presented during the analysis.
The model did not provide an operational bridge explaining how such an EBITDA level could be achieved.
This should have triggered an automatic internal consistency check.
2.3 Failure to incorporate operating constraints
Akenerji’s main natural-gas combined-cycle plant, Erzin, had experienced very low utilization during the relevant period.
The analysis supplied to Gemini included an approximately 14% utilization figure.
Nevertheless, the subsequent valuation did not adequately incorporate this operational constraint and instead used forward EBITDA assumptions that effectively implied a very substantial operational recovery without demonstrating the required electricity prices, natural-gas costs, spark spreads, utilization rates, or production volumes.
A valuation model should not simply assume normalization; it should demonstrate the operational conditions required for normalization.
2.4 Cash-flow constraint failure
The H1 2026 filing showed approximately TRY 733 million of negative operating cash flow.
Despite this, Gemini generated scenarios suggesting substantial deleveraging without first establishing a credible bridge from EBITDA to free cash flow.
This is particularly important given Akenerji’s approximately TRY 25.3 billion net debt.
A company with significant net debt and negative operating cash flow cannot be assigned a credible deleveraging trajectory simply by assuming higher EBITDA.
2.5 Inconsistent price-to-book calculations
Gemini reported both approximately 0.46x and approximately 0.35x price-to-book ratios as if they were current figures.
These cannot both represent the same valuation date and denominator.
The inconsistency was only identified after we explicitly challenged the model.
This illustrates why derived valuation metrics should always carry their underlying inputs:
share price + shares outstanding + equity + reporting date = P/B.
- Other Recurring Problems
Beyond AKENR, we have observed:
Contradictory net-debt definitions
The model has cited different net-debt figures from company reports and brokerage/third-party sources without clearly reconciling differences in:
derivatives,
lease liabilities,
restricted cash,
FX positions,
short-term investments,
debt classification.
A model should state the definition being used before calculating leverage.
Historical operational inputs presented as current
The model has also reused historical spark spreads and price-cap assumptions—for example, a historical -TRY 558/MWh spread-related figure—as if they represented current operating economics.
Historical data can be useful, but it must be explicitly labelled by date and cannot silently become a current assumption.
- What Worked
It is important to note that the problem is not primarily arithmetic capability.
When provided with verified inputs, Gemini generally performed the underlying calculations correctly.
For example, once the inputs were explicitly supplied, calculations involving:
market capitalization,
enterprise value,
approximately 45.6% relationships,
approximately 0.35x P/B,
were mathematically correct.
The principal weaknesses appear to be:
source grounding, period discipline, financial-context verification, self-consistency checking, and adherence to assigned analytical constraints.
- Recommended Technical Improvements
We would strongly recommend consideration of the following safeguards for financial-domain applications.
5.1 Mandatory source metadata for financial figures
Every numerical financial claim should carry:
source,
document/report title,
reporting date,
period,
page/note number where applicable,
whether it is reported or model-derived.
If no source can be verified, the model should explicitly state:
“Source not verified.”
5.2 Strict source hierarchy
The model should prioritize:
KAP/regulatory filing → company financial statements → company investor presentation → audited/reviewed report → reputable third-party research → third-party databases.
A third-party database figure should never silently replace a figure from the primary filing.
5.3 Deterministic financial-math validation
Before presenting a valuation conclusion, the system should automatically check:
EV = market cap + net debt,
EV/EBITDA consistency,
P/B consistency,
debt/equity consistency,
FCF reconciliation,
implied EBITDA behind a stated multiple,
implied market capitalization behind a target price,
period consistency between numerator and denominator.
For example, if an EV of TRY 30.9 billion and a 5x EV/EBITDA multiple are used, the implied EBITDA should automatically be checked as approximately TRY 6.2 billion before the output is presented.
If that result is inconsistent with the operational assumptions, the model should flag the inconsistency rather than proceed.
5.4 Explicit distinction between reported figures and assumptions
The model should clearly distinguish:
Reported:
Figures appearing in the company’s filings.
Derived:
Mathematical calculations based on reported figures.
Assumption:
A user/model input that has not been independently verified.
Scenario:
A conditional outcome based on explicit assumptions.
Forecast:
A projection supported by a defined methodology.
These categories should never be silently substituted for one another.
5.5 Preservation of role constraints
If a user assigns the model a specific role such as:
“Synthesize verified outputs only; do not introduce new numerical assumptions.”
that constraint should persist across subsequent turns.
The model should not claim that a correction has been implemented unless the following output demonstrably reflects the correction.
5.6 Premise verification
Before answering a quantitative question, the model should verify that the question’s premises are themselves correct.
If the premise is incorrect—for example, confusing a hydroelectric plant maintenance shutdown with maintenance at a natural-gas combined-cycle plant—the model should correct the premise before performing the requested analysis.
- Request for Review
We understand that financial analysis is a complex domain and that no model will be error-free.
Our concern is specifically the repeated nature of these errors after explicit correction, particularly when the errors involve numerical claims that appear authoritative but lack verifiable source support.
Given that these outputs can influence investment decisions and portfolio allocation, we believe the issues warrant review by Gemini’s:
financial-domain evaluation team,
factuality/retrieval team,
model alignment team,
and safety/reliability team.
We are willing to provide anonymized transcripts, the relevant KAP filings, and our verification notes if they would be useful for evaluation or model improvement.
Primary filing used for verification: Akenerji KAP Interim Activity Report, 13 August 2026
Thank you for taking this feedback seriously. We hope these examples can contribute to improving Gemini’s reliability, source discipline, and financial reasoning in high-stakes analytical contexts.
Kind regards,
Ömer Sevim
AI Investment Portfolio / AIST